925 81 10MB
Pages 141 Page size 499.92 x 498.72 pts Year 2008
.
rl
•
r
Christopher G. Small
Springer Series in Statistics Andersen/Borgan/Gill/Keiding: Statistical Models Based on Counting Processes. Andrewslllertberg: Data: A Collection of Problems from Many Fields for the Student and Research Worker. Anscombe: Computing in Statistical Science through APL. Berger: Statistical Decision Theory and Bayesian Analysis, 2nd edition. Bolfarinel'Zacks: Prediction Theory for Finite Populations. Bremaud: Point Processes and Queues: Martingale Dynamics. Brockwell/Davis: Time Series: Theory and Methods, 2nd edition. DaleylVereJones: An Introduction to the Theory of Point Processes. Drhaparidze: Parameter Estimation and Hypothesis Testing in Spectral Analysis of Stationary Time Series. FahrmeirlTutz: Multivariate Statistical Modelling Based on Generalized Linear Models. Farrell: Multivariate Calculation. Federer: Statistical Design and Analysis for Intercropping Experiments. Fienberg/HoagliniKruskal/Tanur (Eds.): A Statistical Model: Frederick Mosteller's Contributions to Statistics, Science and Public Policy. Fisher/Sen: The Collected Works of Wassily Hoeffding. Good: Permutation Tests: A Practical Guide to Resampling Methods for Testing Hypotheses. GoodmaniKruskal: Measures of Association for Cross Classifications. Grandell: Aspects of Risk Theory. Haberman: Advanced Statistics, Volume I: Description of Populations. Hall: The Bootstrap and Edgeworth Expansion. Hdrdle: Smoothing Techniques: With Implementation in S. Hartigan: Bayes Theory. Heyer: Theory of Statistical Experiments. Huet/Bouvier/GruetlJolivet: Statistical Tools for Nonlinear Regression: A Practical Guide with SPLUS Examples. Jolliffe: Principal Component Analysis. Kolenlllrennan: Test Equating: Methods and Practices. Kotzllohnson (Eds.): Breakthroughs in Statistics Volume I. Kotzllohnson (Eds.): Breakthroughs in Statistics Volume II. Kres: Statistical Tables for Multivariate Analysis. Le Cam: Asymptotic Methods in Statistical Decision Theory. Le Cam/Yang: Asymptotics in Statistics: Some Basic Concepts. Longford: Models for Uncertainty in Educational Testing. Manoukian: Modem Concepts and Theorems of Mathematical Statistics. Miller, Jr.: Simultaneous Statistical Inference, 2nd edition. Mosteller/Wallace: Applied Bayesian and Classical Inference: The Case of The Federalist Papers.
The Statistical Theory of Shape With 46 Illustrations
{comtnued after indu)
"
Springer
Christopher O. Small Department of Statistics and ActuarialScience University of Waterloo Waterloo, Ontario Canada N2L 301 [email protected]
Preface
I.
Library of Congress CataloginginPublication Data Small, Christopher G. The statistical theory of shape I Christopher G. Small p. em.  (Springer series in statistics) Includes bibliographical references and index. ISBN 0387947299 (hard: alk. paper) 1. Shape theory (Topology)Statistical methods. I. Title. II. Series. QA612.7.S58 1996 514dc20 9613587 Printed on acidfree paper. © 1996 SpringerVerlag New York, Inc. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (SpringerVerlag New York, Inc., 175 Fifth Avenue, New York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use of general descriptive names, trade names, trademarks, etc., in this publication, even if the former are not especially identified, is not to be taken as a sign that such names, as understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone. Production managed by Bill Imbomoni; manufacturing supervised by Jeffrey Taub. Cameraready copy created from the author's LaTex files. Printed and bound by BraunBrumfield, Inc., Ann Arbor, MI. Printed in the United States of America. 987654321 ISBN 0387947299 SpringerVerlag New York Berlin Heidelberg
SPIN 10524357
I
In general terms, the shape of an object, data set, or image can be defined as the total of all information that is invariant under translations, rotations, and isotropic rescalings. Thus two objects can be said to have the same shape if they are similar in the sense of Euclidean geometry. For example, all equilateral triangles have the same shape, and so do all cubes. In applications, bodies rarely have exactly the same shape within measurement error. In such cases the variation in shape can often be the subject of statistical analysis. The last decade has seen a considerable growth in interest in the statistical theory of shape. This has been the result of a synthesis of a number of different areas and a recognition that there is considerable common ground among these areas in their study of shape variation. Despite this synthesis of disciplines, there are several different schools of statistical shape analysis. One of these, the Kendall school of shape analysis, uses a variety of mathematical tools from differential geometry and probability, and is the subject of this book. The book does not assume a particularly strong background by the reader in these subjects, and so a brief introduction is provided to each of these topics. Anyone who is unfamiliar with this material is advised to consult a more complete reference. As the literature on these subjects is vast, the introductory sections can be used as a brief guide to the literature. A few comments should be made about the numbering of figures and propositions. Figures are numbered in order within chapters. Thus Figure 2.3 is the third figure to be found in Chapter 3. Propositions, lemmas, corollaries, and definitions are numbered consecutively within each section. Thus Proposition 2.6.3 is the third result (whether proposition, lemma,
vi
Preface
etc.) within Section 2.6. . Chapter 1 is the basic introductory chapter for the rest of the book. Many of the ideas that are developed in greater detail later are touched upon briefly in this first chapter. Chapter 2 is essentially a review of some basic tools from differential geometry and groups of transformations of Euclidean space. The reader who is familiar with these methods can skim this material for the notation that will be used throughout the rest of the book, and proceed to the next chapter. Chapter 3, which describes various ways of representing shapes on manifolds, is pivotal for all later material, and leads into Chapters 4 and 5, where a stochastic theory is developed on the shape manifolds. Chapter 6 has a collection of applications that are rather loosely bound together by the theme of this book. This book would not have been written without the support of a number of people. Thanks are due to Martin Gilchrist at Springer, who approached me about writing a book on shape. Thanks must go to John Kimmel, also of Springer, whose timely and supportive responses to all my questions made the job of writing much easier. To Springer's production staff and my copyeditor, David Kramer, I offer my sincere thanks. Whenever I had a problem in computing I turned to my colleague Michael Lewis, whose assistance was invaluable. Some of the betterlooking pictures in this book are there through his help. Thanks also go to David Kendall, for his inspiration and valued support over the years. I first began to work on shape theory when I started my Ph.D. under David Kendall's supervision in 1978. What is good in this book is largely due to him. What is bad is my responsibility alone! Thanks also to my colleagues Huiling Le and Colin Goodall for their excellent advice on the subject, and to Fred Bookstein for his insights and energy. Zejiang Yang was also very helpful in catching a number of errors in the manuscript. I could not conclude this checklist of indebtedness without acknowledging the support of my wife Kristin Lord, who put up with the long hours I spent working on the manuscript. Kristin was also instrumental in bringing the Mt. Tom dinosaur data set to my attention. Christopher G. Small University of Waterloo June 1996
Contents
Preface 1 Introduction 1.1 Background of Shape Theory 1.2 Principles of Allometry. . . . 1.3 Defining and Comparing Shapes 1.4 A Few More Examples . . . . . . 1.4.1 A Simple Example in One Dimension 1.4.2 Dinosaur Trackways From Mt. Tom, Massachusetts. 1.4.3 Bronze Age Post Mold Configurations in England 1.5 The Problem of Homology . 1.6 Notes .. 1.7 Problems 2
Background Concepts and Definitions 2.1 Transformations on Euclidean Space 2.1.1 Properties of Sets. . . . . . . 2.1.2 Affine Transformations. . . . 2.1.3 Orthogonal Transformations. 2.1.4 Unitary Transformations. . . 2.1.5 Singular Value Decompositions 2.1.6 Inner Products . . . . . . . . . 2.1.7 Groups of Transformations " 2.1.8 Euclidean Motions and Isometries 2.1.9 Similarity Transformations and the Shape of Sets. .2.2 Differential Geometry .. . . . . . . . . . . . . 2.2.1 Homeomorphisms and Diffeomorphisms 2.2.2 Topological Spaces . . . . . . . . . . . 2.2.3 Introduction to Manifolds . . . . . . . 2.2.4 Topological and Differential Manifolds 2.2.5 An Introduction to Tangent Vectors .
v 1 1 4 6 14 14 16 20 24 26 27 29 29 29 29 30 30 31 32 33 34 34 36 36 37 38 39 41
viii
Contents
Contents 2.2.6 Tangent Vectors and Tangent Spaces . . . . 2.2.7 Metric Tensors and Riemannian Manifolds. 2.2.8 Geodesic Paths and Geodesic Distance. 2.2.9 Affine Connections . 2.2.10 Example ' . 2.2.11 New Manifolds From Old: Product Manifolds 2.2.12 New Manifolds From Old: Submanifolds .. 2.2.13 Derivatives of Functions between Manifolds 2.2.14 Example: The Sphere . 2.2.15 Example: Real Projective Spaces . . . 2.2.16 Example: Complex Projective Spaces. 2.2.17 Example: Hyperbolic Half Spaces. 2.3 Notes 2.4 Problems
3 Shape Spaces 3.1 The Sphere of Triangle Shapes . 3.2 Complex Projective Spaces of Shapes . 3.3 Landmarks in Three and Higher Dimensions . 3.3.1 Introduction . 3.3.2 Riemannian Submersions . 3.4 Principal Coordinate Analysis . . . . . . . . . 3.5 An Application of Principal Coordinate Analysis 3.6 Hyperbolic Geometries for Shapes . 3.6.1 Singular Values and the Poincare Plane . 3.6.2 A Generalization into Higher Dimensions 3.6.3 Geodesic Distance in UT(2). . . . . . 3.6.4 The Geometry of Tetrahedral Shapes. 3.7 Local Analysis of Shape Variation . . . . . . 3.7.1 ThinPlate Splines . . .. . . . . . . . 3.7.2 Local Anisotropy of Nonlinear Transformations 3.7.3 Another Measure of Local Shape Variation 3.8 Notes 3.9 Problems . 4 Some Stochastic Geometry 4.1 Probability Theory on Manifolds 4.1.1 Sample Spaces and SigmaFields 4.1.2 Probabilities . 4.1.3 Statistics on Manifolds . 4.1.4 Induced Distributions on Manifolds . 4.1.5 Random Vectors and Distribution Functions . 4.1.6 Stochastic Independence .. 4.1.7 Mathematical Expectation . 4.2 The Geometric Measure . . . . . .
44 47 48 50 51 51 52 52 54 55 59 62 66 66
4.3
.
"", 4.4
4.5
4.6
69 69 77 79 79 84 87 92 95 95 99 104 105 106 106 110 112
114 114 117
117 117 118 118 119 120 121 121 121
4.7
4,8 4.9
4.2.1 Example: Surface Area on Spheres . . . . . . . 4.2.2 Example: Volume in Hyperbolic Half Speces., Transformations of Statistics . . : . . 4,3.1 Jacobians of Diffeomorphisms . 4.3.2 Change of Variables Formulas. Invariance and Isometries . . . . . . . 4.4.1 Example: Isometriesof Spheres 4A.2 Example: Isometries of Real Projective Spaces" 4.4.3 Example: Isometries of Complex Projective Spaces Normal Statistics on Manifolds . . .' . . . 4.5.1 Multivariate Normal Distributions . .. 4.5.2 Helmert Transformations 4.5.3 Projected Normal Statistics on Spheres Binomial and Poisson Processes . . . . . . . 4.6.1 Uniform Distributions on Open Sets . 4.6.2 Binomial Processes. . . . . . . . . . . 4.6.3 Example: Binomial Processes of Lines 4.6.4 Poisson Processes.. . . . . . . . . . . Poisson Processes in Euclidean Spaces . . . . 4.7.1 Nearest Neighbors in a Poisson Process 4.7.2 The Nonsphericity Property of thePP . 4.7.3 The Delaunay Tessellation. . . . . . . . 4.7.4 PreSizeandShape Distribution of Delaunay Simplexes Notes Problems . . . .
ix
123 123 124 124 124 125 127 127 129 130 130 130 131 134 134 134 135 137 139 139 140 141 143 145 147
5 Distributions of Random Shapes 5.1 Landmarks from the Spherical Normal: lID Case 5.2 Shape Densities under Affine Transformations . . 5.2.1 Introduction 5.2.2 Shape Density for the Elliptical Normal Distribution 5.2.3 Broadbent Factors and Collinear Shapes. 5.3 Tools for the Ley Hunter. . . . . . . . . . . . . . . . . 5.4 Independent Uniformly Distributed Landmarks . . . . 5.5 Landmarks from the Spherical Normal: NonIll) Case 5.6 The PoissonDelaunay Shape Distribution 5.7 Notes .. 5.8 Problems
149 149 152 152 154 156 158 162 163 167 170 171
6 Some Examples of Shape Analysis 6.1 Introduction.......... 6.2 Mt. Tom Dinosaur Trackways 6.2.1 Orientation Analysis 6.2.2 Scale Analysis . . . .
173 173 173 174 176
x
Contents 178 6.2.3 Shape Analysis ............ 180 6.2.4 Fitting the MardiaDryden Density. 182 6.3 Shape Analysis of Post Mold Data ..... 182 6.3.1 A Few General Remarks . . . . . . . 184 The Number of Patterns in a Poisson Process 6.3.2 187 Post Molds An Annular Coverage Criterion for 6.3.3 6.4 Case Studies: Aldermaston Wharf and South Lodge Camp . 190 190 6.4.1 Scale Analysis 191 6.4.2 Shape Analysis 193 6.4.3 Conclusions . . 193 Automated Homology 6.5 193 6.5.1 Introduction 194 6.5.2 Automated Block Homology . 197 6.5.3 An Application to Three Brooches 199 6.6 Notes .................... 199 6.6.1 Anthropology, Archeology, and Paleontology ....... . . 199 and Medical Sciences Biology 6.6.2 199 6.6.3 Earth and Space Sciences ......... . . 199 6.6.4 Geometric Probability and Stochastic Geometry .. 199 6.6.5 Industrial Statistics . . . . . . . . . . . . 200 6.6.6 Mathematical Statistics and Multivariate Analysis 6.6.7 Pattern Recognition, Computer Vision, and Image 200 Processing. . . . . . . . . . . . . 200 6.6.8 Stereology and Microscopy . . . . 200 6.6.9 Topics on Groups and Invariance Bibliography
201
Index
217
1 Introduction
~

1.1
Background of Shape Theory
In 1977, David Kendall published a brief note [87] in which he introduced a new representation of shapes as elements of complex projective spaces. The result stated in the paper was unusual: under an appropriate random clock, the shape of a set of independent particles diffusing according to a Brownian motion law could be regarded as a Brownian motion on complex projective space. Many statisticians, who knew little about complex projective spaces and who did not work on diffusion processes, did not see immediate applications to their own work. However, in a sequence of talks at conferences around the world, David Kendall continued to expound on his theory, with some applications to problems in archeology. Presented with great clarity and with excellent graphics, these talks gradually generated wider interest. It was not until 1984 that the full details of the theory were published [90]. At that point it became clear that Kendall's theory of shape was of great elegance and contained some interesting areas of research for both the probabilist and the statistician. The full range of possible applications became much clearer when David Kendall was invited to be a discussant for a paper by Fred Bookstein [19] in the journal Statistical Science. Kendall and Bookstein, it turned out, had been thinking along the same lines, namely that shapes could be represented on manifolds. There were some intriguing differences. Whereas Kendall represented the shapes of triangles in the plane as points on a sphere, a space of positive curvature, a suggestion of Bookstein represented
2
1.1 Background of Shape Theory
1. Introduction
the shapes of those triangles as points on a Poincare half plane, a spaceof negative curvature. Perhaps more important were the different applications each researcher emphasized. Kendall's applications were in the archeological and astronomical sciences, and studied the shapes of random sets of points, such as are to be found in a Poisson scattering. Bookstein's applications were in the biological and medical sciences, and drew on the tradition of researchers such as D' Arcy Wentworth Thompson, Julian Huxley, and later researchers in allometry and multivariate morphometries. For Bookstein and his colleagues, the points under consideration were biologically active sites on organisms called landmarks. At present, we can speak of both Kendall and Bookstein schools of shape analysis. It is within this context that this book is written. The primary theme of the book will be the representation of shapes on differential manifolds, and the statistical consequences of this idea. The emphasis will be more toward the Kendall school, where the differential geometry of shape analysis is more developed. However, we shall frequently compare this with some of the work of Bookstein and others, insofar as this is relevant to our goal. In tracing the history of methods that have produced this statistical theory of shape, it is quickly apparent that a great variety of past work is responsible for its development. It is difficult to imagine a time in history when people have not been fascinated by shapes. Our visual fine arts, such as painting and sculpture, have appeal across cultures and illustrate the universality of shapes or forms. As D'Arcy Thompson pointed out in his pioneering book 'On Growth and Form [172], there is an important relationship between the form or shape of a biological structure and its function. Thus the study of shape is also the study of function. For example, the mathematical constraint that a: body have minimum surface area for a given volume requires that it be roughly spherical in shape. This result is known in mathematics as an isoperimetric inequality, and can be used to explain why an organism that seeks to minimize its boundary with an external environment, for heat conservation or defense, will often have a simple spherical curvature. On the other hand, if the boundary of an organism is required to be permeable to allow oxygen or food to flow across it, then such a minimization of surface area would be inappropriate. One would expect the surface area in this case to be roughly proportional to the volume of the organism. However, an organism cannot grow while maintaining the same shape and continuing to have a constant ratio between its volume and its external surface area. In this case, growth usually involves a change of shape, possibly through the introduction of a highly convoluted boundary. The geometric structure of lung tissue is a case in point. Recent developments in the theory of fractal shapes have shown that the boundary of a threedimensional structure need not scale upwards as the square of its length or diameter. In fact, a highly convoluted surface can be thought of as an approximation to a fractal.
:1
3
Sometimes the relationship between shape and function is of a more contrived nature. For example, the amphorae used in antiquity had a variety of forms. The particular shapes of amphorae were guides to the nature of the contents. This relationship continues down to the present day: nobody need confuse the contents of a bottle of white wine with the contents of a bottle of whiskey, as shape tells all. Much of statistical theory has been dedicated to the estimation of location and scale parameters. As the statistical theory of shape is concerned with aspects of the data. that remain after location and scale information are discounted, statistical shape concepts have not been as prominent as the theory of inference for location and scale. In 1934, R. A. Fisher [57] introduced the concept of the configuration of a univariate sample. This concept is equivalent to the formal definition of shape for dimension one that we shall develop in this book. In 1939, E. J. G. Pitman [134] developed the theory of minimum variance equivariant estimation of location and scale parameters, and in so doing illustrated the importance of conditioning on invariant statistics in the construction of best equivariant estimators for location and scale. The idea of a shape statistic as a maximal invariant under location and scale transformation can be seen in this work, although the shape .statistics play an ancillary role to the estimation of parameters associated with location and scale transformations. The extension of the concept of invariance to multivariate data is straightforward. However, it is in the psychometric literature that statistical tools were first developed for the comparison of the shapes of data sets. The roots of.Procrustes analysis can be traced to Mosier [123], and then through the work of Sibson [150, 151] and Gower [75]. In comparing the differences in shape between two data sets, Procrustes analysis proceeds by transforming one data set to try to match the other. The transformations allowed in' a standard analysis include shifts in location, scale changes, and rotations. Together, these transformations are called similarity transformations or shapepreserving transformations. When a transformation of one of the data sets has been found to most nearly match the other, the sum of squared differences of the coordinates between them is called the Procrustean distance between the two data sets. We shall see that this concept is closely related to the natural measure on distance between shapes that we shall consider in Section 1.3. Another line of research that has contributed to the statistical theory of shape is to be found in the field of geometric probability and stochastic geometry. It is here that we see geometric objects, such as points, lines, and convex sets, as the basic data for the statistician. The set of outcomes of a random experiment can often be represented as a region in space whose volume, or pdimensional content, can be ascertained. Within this region is to be found some subset E corresponding to an event. According to one definition, the probability P(E) of this event is the ratio of the pdimensional content of this subset to that of the entire region. Such a
4
1.
definition is problematic for certain applications, and leads to paradoxes such as that of Bertrand involving random lines. For this reason, the modern theory of geometric probability makes use of invariance of probabilities under Euclidean motions as a more fundamental notion for calculating the probability of geometric events. That is, a probability measure can be said to be geometric if it assigns equal probability that a random geometric object such as a point or line will hit congruent sets. In 1980, David Kendall and his son Wilfrid Kendall [95] proposed the use of techniques from geometric probability to examine the hypothetical alignments of megalithic stones from Land's End in Cornwall. This data set of fiftytwo sites at Land's End was originally investigated by Alfred Watkins [177], who advanced the theory that megalithic cultures had deliberately placed standing stones along straight lines known as ley lines. The folklore around the existence and interpretation of such lines is quite extensive despite the patchy evidence for the existence of ley lines. Kendall and Kendall [95] followed the approach of Simon Broadbent [33J by calculating the expected number of approximately collinear triplets of points if the megalithic sites had been positioned at random. As a triangle can be called approximately flat (or eblunt in their terminology) if its maximum internal angle is within tolerance € of a straight angle, Kendall and Kendall were naturally drawn to the examination of the distribution of angles in a random triangle, and thereby to the concept of an induced marginal distribution on a space of triangle shapes. The paper by David Kendall [90] in 1984 was seminal for the development of the geometry and distribution theory of shape space. A key result of this paper was that the induced distribution of shape for a set of independent identically distributed bivariate normal points is uniform on the shape space when the covariance matrix is a multiple of the identity. The univariate version of this result also holds, although it is of much older vintage than the bivariate result. The work of Dryden and Mardia [53, 116, 117] generalized this work to the shapes of points from bivariate normal distributions having different means, and set the stage for the distribution theory to be tied in to the work on shape analysis developed in allometry, to which we now turn.
1.2
1.2 Principles of Allometry
Introduction
Principles of Allometry
Allometry can be defined as the study of the relationship between size and shape. If we take a set of measurements of distances between points on a body, then a size variable can be regarded as a summary of the overall scale of these measurements. For example, the arithmetic mean and the geometric mean of a set of distances are both size variables. Size variables are required to be homogeneous functions of the set of distances. This means that if all measurements are increased or decreased by a common scale
5
factor, then the size variable is itself increased or decreased by that same factor. If we standardize the distances by scaling them to have unit size variable, then the resulting ratios of dimensions are called shape variables. Many of the key insights into the growth allometry of biological organisms were first outlined by Julian Huxley [85]. Allometry studies shape differences by taking ratios of dimensions of objects. As much of statistics is linear in nature, it is natural to take logarithms of the dimensions of objects and plot these logarithmic coordinates on a graph. Now, two objects of different size but common shape will have their dimensions in the same ratio. Therefore the shape statistics can be associated with differences between the logarithmic dimensions. For example, suppose we consider how an organism changes shape as it matures and grows with age. Let Xt and Yt be two recorded dimensions of the organism at age t, so that Yt/Xt is a partial description of the shape of the organism. Now, if all parts of the organism grow at a constant rate a as it matures, then growth will be exponential in nature, and we will have the formulas Xt = Xo exp(at)
Yt
= Yo exp(at)
(1.1)
Thus
(1.2) which does not involve the age t of the organism. So the logarithmic coordinates (log Xtjl log YtJ, when plotted at different ages tj, will all lie on a line of slope +1, which corresponds to constancy of shape. On the other hand, if these coordinates do not all lie on a line of slope +1, then we can deduce that there is some variation in shape between different ages. However, if Xt grows at a constant rate a and Yt grows at rate {3 =I a, then these logarithmic coordinates plotted at different ages will still lie on a straight line. In this case, the slope of the line will differ from unity. This fact, namely that the logarithms of size variables lie on straight lines, is one of the basic empirical principles of allometry. This empirical principle has a theoretical foundation in a model that presupposes exponential growth at varying rates in different parts of an organism. In turn, this variation in the growth rate explains some of the variation in the shape of an organism as it matures. It should be noted that the size variables need not be linear in nature in order that their logarithms lie on straight lines. We can extend from comparing distances of bodies or organisms to more general size variables such as surface area or volume, and we will still keep a linear functional relationship between their logarithmic coordinates if growth is exponential. The effect of using an area, say, rather than a length for Yt is to scale the slope by a factor of two in the plot of log(xtJ and 10g(YtJ. The analysis is seen to be statistical in nature when we reflect on the fact that measurement error and a slight unevenness of growth are to be expected under normal circumstances. Therefore, even when the model
6
assumptions are correct, we would.notexpectshepoints to lie ona perfect straight line. Statistical tools such as principal components analysis can be used to draw a line through the data. This is equivalent to fitting a bivariate normal distribution to the scatter plot of points (log Xtj' log Ytj) and finding the principal axis through the elliptical contours of the normal density. At first sight, the extension from two size variables Xt and Yt to several would seem to be easy. While the linear statistical analysis ofmultivariate data through principal components is straightforward, the extension is problematic because the assumption of multivariate normality is quite stringent. In typical data sets, the set of size variables such as lengths have complicated nonlinear relationships among them. For .example, if we were to record a set of 21 interpoint distances between 7 points on a twodimensional image, we would only have 11 degrees of freedom among the 21 distances. The particular restrictions on these variables would be complicated and nonlinear, and would make modeling of their logarithms using normal assumptions difficult. It is at this point that the techniques of Procrustes analysis provide an avenue of escape from these difficulties. The problems that arise in taking ratios of size variables point us toward nonlinear mathematics and towards a theory of shape based upon configurations of points rather than ratios of size variables. This theory of shape will be the central topic of the book.
1.3
1.3 Defining and Comparing Shapes
1. Introduction
Defining and Comparing Shapes
When all information in a data set about its location, scale, and orientation is removed, the information that remains is called the shape of the data. Alternatively, we can say that two data sets have the same shape if a combination of a rigid motion and rescaling of one of the data sets will make it coincide with the other. In geometry, two figures that have the same shape are said to be similar. For example, two triangles will be similar provided their corresponding internal angles are equal. To investigate the concept of shape more carefully, consider Figure 1.1, which shows three examples of side views of Iron Age brooches from a cemetery excavated at modernday Miinsingen, inSwitzerland. As these brooches can be ordered chronologically from the layout of the cemetery, it is natural to consider how the shapes of the brooches developed over time. These three brooches represent only a fraction of the total data from the cemetery but will serve the purpose here of illustrating some basic principles of shape analysis. Let us suppose for the moment that we are given these pictures as our primary data. How can we analyze the differences in shapes of the three brooches? A first step in such an analysis might be to construct a finite
c:2b ~ 4
Brooch 1
3 '
~ 2
7
~ 3
3
~ ~ 4 3
4
Brooch 2
Brooch 3
FIGURE 1.1. Three Iron Age brooches. Prom each of the images, four landmarks are chosen at locations coinciding with features in the brooches. The landmarks correspond in a natural sense, so that landmarks in different images marking corresponding features are labeledin a similar fashion. The shape analysis proceeds by eliminating information in each of the configurations about location, scale and orientation. The brooches are adapted from Hodson, Sneath, and Doran, Biometrika 53 {1966}, p. 315, by kind permission of Biometrika Trustees.
dimensional representation of some of the important geometric information from each picture. For example, we could construct a set of points Xl, X2, ••• , X n lying on each figure such that the locations of these points coincide with important features. On different bodies or figures, sites used for summarizing or encoding of geometric information are called landmarks. For example, on the human face, the positions of the eyes and other features can be used as landmarks to analyze the shape of the face. For our purposes, landmarks will be defined as points chosen from an image or object to mark the location of important features and to give a partial geometric description of the image or object.
Normally, we think of the features of a twodimensional image as lying in a very highdimensional space, or, in an idealized sense, in an infinitedimensional space. If we keep this in mind, then we recognize that there is inevitably some loss of information in encoding pictures with a relatively small number of landmarks. Nevertheless, small numbers of landmarks can provide the basis for comparisons of important shape differences. Just as a small number of landmarks within a city might help us find our way around by identifying features of the city, so the landmarks chosen to summarize a figure can be regarded as identifying its important geometric features. Let us consider .how a set of four landmarks can be constructed for each of the three images. The centers of the coiled springs on the right of each figure represent corresponding points, and similarly, the leftmost points at which the curvature is sharpest also correspond. For each brooch, let Xl , and X2 be these two points respectively. Additionally, let X3 be the upper point on each brooch where the left piece bends back and fastens to form a
8
1. Introduction
1.3 Defining and Comparing Shapes
loop. Finally, we can choose the fourth landmark X4 to be the lower bend on the loop. (This is the point of high curvature in the loop where the pin is secured.) This locates four landmarks for each figure. More generally, n landmarks can be chosen so that the vector (Xl, ..., x n ) , which lies in (R 2)n, provides a 2ndimensional summary of some of the major geometric characteristics of the brooch, including location, orientation, scale, and shape information. To perform a shape analysis on these landmarks, we must determine the class of all functions of the vector (Xl, ..., x n ) that measure its shape. This involves the elimination of information in (Xl, ..., x n ) that describes the location, scale, or orientation of the landmarks. The location and scale statistics of a set of points are perhaps best known to statisticians because they can be described by standard statistical tools. For example, the location of a data set (Xl, ..., X n) can be described by its sample mean, or centroid, given by 1 X
n
(1.3)
=  LXj
n
j=l
In addition, the size or scale of our configuration of landmarks can be described by a variety of statistics. Let us choose coordinates for each of the landmarks so that Xj = (Xj!, Xj2) for j = 1, ... , n, and x = (Xl, X2). The column vectors of residuals about the means are
9
is a natural measure of the size of the set of landmarks because it is independent of the orientation of the Cartesian coordinate system. The usual way to eliminate location and size information in data is by standardization, which is a combination of a location shift and a rescaling so that the data set has centroid X at the origin in R 2 and the matrix I' is standardized to have trace equal to one. For our example, the standardized data set becomes (1.8)
A caveat must be mentioned here. In order for this representation to be meaningful, the landmarks Xl, ..., X n must not all be coincident. This presents no problem for our application to brooches. In general, a set of landmarks that are all coincident will be said to have indeterminate shape. Henceforth, we shall assume that this degeneracy does not arise. Note, however, that we do not exclude cases in which some but not all of the landmarks are coincident. We shall refer to the vector T defined in (1.8) as the preshape of the landmarks. While this terminology is not particularly descriptive, it does emphasize the order in which the reduction to shape progresses. The preshape 2 T lies in a constrained subset of the original Euclidean space (R )n. This subset can be represented by the intersection of the (n  2)dimensional subspace n
(1.4)
(1.9)
with the unit sphere and
n
LllXjW =
I}
(1.10)
j=l
(1.5)
The intersection (1.11)
X n2 
X2
Then the matrix of squared residuals can be written as (1.6)
where (.)T denotes the transpose operation. The trace of I', given by n
LllXj xW j=l
(1.7)
is a (2n  3)dimensional sphere within the ambient Euclidean space R 2n . . A subscripted star is included as a gentle reminder to ourselves that this (2n  3)dimensional sphere is not the usual unit sphere embedded in R 2 n  2 . We shall refer to this sphere as the preshape space or the sphere of preshapes. At the next stage of our analysis, we must eliminate the information about the orientation of the data set, in order that the quantity which remains be a shape statistic. At first glance, the problem of defining and standardizing the orientation of the preshape of the data would seem to be similar to the problems of defining and standardizing the location and
10
1. Introduction
1.3 Defining and Comparing Shapes
scale. However, this is not the case. Some topological problems arise that cannot easily be r e m o v e d . , By the orientation of a set of planar landmarks we intuitively understand the angle made by some axis through the landmarks with respect to some given axis, independent, of the landmarks., For example, we could, use the angle made by a ray from Xl to X2 as the description of the orientation of (Xl> ..., x n ) . While this will be quite satisfactory for the data that we are considering here, it will not suffice for orienting all configurations (Xl, ... , x n ) . Those sets of landmarks for which Xl = X2 cannot be oriented by such a definition. Of course, another definition can be used for these preshapes. However, we would obviously like to do better than this by finding ? single definition that works for all samples. Any angle can be represented as a point on Sl, the unit circle about the origin in R 2 . So the orientation of (Xl, ... , X n ) can be defined as a point 6(XI,""X n ) E s. The process of standardizing the location and scale of (Xl, ..., x n ) does not disturb its orientation. Therefore, we can also refer to the orientation 6(r) of the preshape r. It follows that the orientation of the preshape can be written as a function (1.12)
from the sphere of preshapes into the unit circle of the plane. In addition, it is reasonable to suppose that an ideal orientation function would be a continuous function of its coordinates, so that 6 would be a continuous function on the sphere s;n3. Now suppose that fJ: R 2 + R 2 is a rotation of the plane about the origin. Under the rotation of the landmarks Xj + fJ(Xj), the corresponding preshapes transform as (1.13) This defines a mapping fJ: s;n3 + s;n3 . Note that we abuse terminology slightly by using the symbol fJ to refer to the rotation on R 2 as well as the rotation on s;n3. There is seen to be a simple correspondence between the two that makes the notation convenient. If 6 is an appropriate orientation function, then it should be compatible with the rotations of the plane, so that 6[B(r)] = fJ[6(r)] for any preshape r E s;n3. However, it is here that we get into trouble in attempting to define the function 6. It can be shown that there does not exist a continuous function e: s;n3 + 8 1 that satisfies this property. In order to see this, consider ,the following. The orbit of any preshape r E s;n3 will be the circle O(r)
=
{fJ(r): 0
:s; fJ < 2rr} c s;n3
11
function. If 6 were continuous, then the function (1.15)
would be a retraction of s;n3 onto the circle O(r). That is, 6 16, would be a continuous function onto a subset of s;n3 whose restriction to that subset would be the identity mapping. An argument in algebraic topology using fundamental homotopy groups, which we .omit, shows that this is impossible. Thus we have the following: Proposition 1.3.1. For n > 2 there does not exist a continuous orientation function 6 : s;n3 + s' that is compatible with rotations of the original coordinates (X1,""X n ) in the sense that 6[fJ(r)] = fJ[6(r)] for all r E s;n3.
Proposition 1.3.1 tells us that continuous methods to standardize the orientation of preshapes will fail. That is, we cannot find a single definition that is continuous in the data and simultaneously orients all preshapes. Our original purpose in standardizing the landmarks (Xl,,,,, Xn ) with respect to location, scale, and orientation was to provide a set of coordinates for their shape. Proposition 1.3.1 does not exclude the possibility of our constructing coordinate systems that work for some shapes but not for others. In fact, we must distinguish between representing shapes and constructing shape coordinates. As we shall see in Chapter 3, shapes are naturally represented as points in a shape manifold. However, there will typically be no single coordinate system on that shape manifold that is nondegenerate and that provides coordinates for all points in the manifold. For example, on the Earth's surface, the coordinates of longitude and latitude work perfectly well except at the poles, where the longitude coordinate is redundant. Coordinates with latitude 900 N and different longitudes refer to the same point, namely the north pole. The failure of a single coordinate system to work at all points on the sphere is simply a reflection of the fact that the sphere is not topologically equivalent to any subset of the plane. . Just as we do not identify the sphere with its coordinate system, 'So we should not identify shapes and .shape representations with any particular coordinates used to construct them. As we shall see, the appropriate setting for representing shapes is as an orbit space I;~ of a sphere s;n3. By an orbit space of the sphere we mean a set I;2' of equivalence classes, namely (1.16)
(1.14)
Therefore 6, when restricted to the orbit O(r), would be a 11 correspondence between O(r) and s. Let 6 1 : s + O(r) be the inverse
Two preshapes r1 and r2 will lie in the same equivalence class O(r) provided they have the same shape. If this is the case, there will exist a rotation fJ such that B(ri) = r2.
12
1.
1.3
Introduction
However, this formal definition of E~ as a set of equivalence classes is of little value unless we can compare shapes and obtain some geometric intuition about E~. To do this, we must define a metric on E2" A metric is a mathematical generalization of the concept of Euclidean distance between points. Metrics have certain properties, which are listed in Problem 5 at the end of the chapter. If we think of E2' as a space, then its elements can be regarded as points in that space, for which we seek an appropriate definition of distance. An obvious way to do this is to use a metric between orbits on n  3 • As preshapes can be represented as points on the preshape space this sphere, the distance between two preshapes is the geodesic, or great circle distance, between preshapes. On the earth, the great circle distance is the shortest distance one would have to travel to get from one place to another. This is quite easy to compute for spheres of any dimension. If 71 n  3 then the great circle distance between and 72 are two preshapes on 71 and 72 is given by
8:
8:
(1.17) where < 71,72 > is the inner product between 71 and 72 as vectors in R 2 n . Note that the cos"! function is defined so as to have range [0,7r]. The induced metric on E2' is then defined as (1.18)
where inf A is the infimum function over any set A of real numbers. In more informal language, we can say that the distance between two shapes, or orbits of 8;"3, is the minimum of the distances between all pairs of preshapes lying in the respective orbits. The reader should note that to perform the minimization, it is sufficient to fix (h and minimize over all values of 82 , or vice versa. Problem 5 at the end of the chapter asks the reader to show that formula (1.18) satisfies the properties of a metric. With this metric, the space E2' turns out be be a manifold. In fact, as we shall see in the next chapter, it is an example of a complex projective space. We shall leave the definition of these spaces to Section 2.3 and shall concentrate for the moment on calculating this metric on the shape space E2" To evaluate this metric on E~ we can make use of the algebraic properties of the complex plane. Suppose we consider the landmarks Xl, ... j Xn to be elements ofthe complex plane by identification of the complex numbers C with R 2 • Then Xk can be regarded as a complex quantity by identifying the two coordinates of Xk E R 2 with the real and imaginary components of a complex number. Under this identification, the preshape coordinates Xk 
7k
=
v"L
X
IXk _ k
xj2
(1.19)
for k = 1,2, ... , n can also be regarded as complex quantities, being standardized versions of the original coordinates.
Defining and Comparing Shapes
13
Let 0"1 and 0"2 be two shapes in E2', and let us choose representative preshapes T1 and T2 so that O"j = 0(7j) for j = 1,2. Write 7j
=
(Tj1, 7j2, ..., 7jn)
(1.20)
where rs» is the kth complex standardized coordinate of Tj. Furthermore, let Tjk be the complex conjugate of 7jk. We will go into the details of the mathematics in Example 2.3.16 of the next chapter. For the moment, we shall note that the minimum in formula (1.18) can be found algebraically to be
d(0"1,0"2) = cos
1
(I ~
7lkT;kl)
(1.21)
This is called the Procrustean distance or the Procrustean metric from 0"1 to 0"2. As the argument of the cos"! function is always nonnegative, we note thecurious fact that the maximum Procrustean distance between shapes in E2' is 7r/2 . The reader should also note that the right hand side of this identity does not depend upon the orientation of the preshapes 71 and 72 . A rotation of these preshapes corresponds to multiplying each Tjk by an element of the unit circle in the complex plane. This factors out of the summation and has modulus one. Let us apply this formula to the shape differences of the landmarks of Figure 1.1. An inspection of this figure would suggest that the landmarks of the second and third brooches are closer in shape to each other than they are to the landmarks of the first brooch. It remains to be seen whether the shape analysis from landmarks supports this conclusion. In each of the three images, we have n = 4 landmarks. Let 71, 72, and 73 be the preshapes of the respective configurations of landmarks shown in Figure 1.1, as defined by formula (1.8). Additionally, let 0"1, 0"2, and 0"3 be the respective shapes of these sets of landmarks. Then from formula (1.21), we get d(0"1,0"2) = 0.380, d(0"1,0"3) = 0.308, and d(0"2,0"3) = 0.132. As would be expected, the smallest shape difference is between the second and third brooches. The first brooch can be distinguished from the other two by the fact that its loop is fastened at the top much further to the right than the others. In terms of landmarks, we can see that' X3 and X4 are shifted closer to Xl in the first brooch than is the case for the second and third brooches. The landmark analysis also suggests that the third brooch is slightly closer in shape to the first brooch than is the second. The three brooches that we have considered here for the sake of example are part of a larger set of brooches. In Section 3.7 we shall conduct a shape analysis of the complete set of images. Of course, such conclusions are dependent upon the choice of landmarks on the brooches. Four landmarks are too few to draw more than crude comparisons between the shapes of the brooches. In Chapter 6 we shall consider methods to study the shape variation between the brooches in finer detail.
14
1. Introduction
1.4 A Few More Examples
While the shape metric provides a geometric structure to ~~, we are still left with a considerable difficulty in interpreting and visualizing this space. In Chapter 3, we shall construct some concrete representations of shape spaces. Moreover, before we leap upon such a choice for the geometry of shape space, it is worth bearing in mind that this choice of metric is closely connected to the concept ofa metric between preshapes. However, the great circle distance between preshapes on the sphere s~n3 is a consequence of the standardization technique, namely the rescaling of the original centered landmarks so that tr(r) = 1, where tr( .. ) is the trace function defined in formula (1.7). The conclusions drawn from a shape analysis based upon a metric geometry of shape space will depend in part upon the choice of size variable used to compare shapes. In Chapter 3, we will examine various geometries of shape space and will find some simple representations for special cases.
15
xl
'cJ::LI..lJ;;:=wI
x3
xl
1.4 A Few More Examples 1.4.1 A Simple Example inDne Dimension Throughout this and subsequent chapters, we shall be primarily concerned with the representations of shapes of landmarks in dimension two or above. However, before proceeding to that material, it is useful to consider what happens with landmarks that lie along a line. First and foremost, we should note that one dimensional configurations of landmarks cannot be rotated. Therefore, the preshapes of such configurations of landmarks can be identified with their shapes, there being nothing more to remove upon reduction to the preshape. This makes the representation of shapes in dimension one a very easy thing to do. Preshapes lie naturally on a sphere. We have seen this, in particular, for landmarks in the plane. However, it remains true for landmarks in any dimension. If we have n::::: 3 landmarks along a line, then the preshape
xl
xl T
, ... ,
(1.22)
of these n landmarks will lie in a sphere (1.23)
of dimension n  2. A sphere of dimension one is, of course, a circle. Even three landmarks along a line can sometimes be used to make basic shape comparisons. Consider for example Figure 1.2, which shows the profiles of four skulls. Also plotted over each of the skulls is a set of three landmarks, chosen according to a landmark selection method proposed by
FIGURE 1.2. Side view of skulls. Prom top to bottom: modern human, Neanderthal, australopithecine, chimpanzee. The skull profiles are redrawn from Figure 3.53 of [131].
16
1.4 A Few More Examples
1. Introduction
Michael Lewis of the University of Waterloo. Upon examination, the four skulls are seen to vary particularly in the ratio of the size of the cranium to the size of the jaw. In the human skull this ratio is the largest, while it is smallest for the chimpanzee. The landmarks Xl, x2, and X3 capture some of this variation because the craniumtojaw ratio is proportional to the ratio of the distances from Xl to X2 and from X2 to X3. As Xl, X2, and X3 lie along a line in each picture, we can put some coordinates along each line and consider (Xl, X2, X3) to be a vector in R 3 . The preshape r of such a vector will then be an element of the unit circle s'. Figure 1.3 shows the preshapes of these four configurations of three points plotted on a circle. The reader may be surprised by the small amount of arc length enclosed within the range of the four preshape points in Figure 1.3. This is quite typical of landmarks chosen on biological organisms. Usually, the amount of variation of landmark coordinates between images is small compared to the distances between the landmarks within an image. A small arc of a circle can be approximated by a line segment. So it is tempting to approximate the positions of preshapes on the circle in Figure 1.3 by a similar configuration along a straight line. Such an approximation is called a tangent approximation, and works quite well for many biological data sets. More generally, however, configurations of points on a circle cannot be approximated by a configuration of points along a line without major distortion of the interpoint distances. Similarly, a configuration of points on a shape space such as I:~ cannot be approximated by a multivariate configuration in R 2n  4 without distorting the interpoint Procrustean distances. So it is fortunate when such a tangent approximation is possible, because it permits the researcher to apply the large collection of multivariate statistical techniques designed for data in Euclidean space. In general, the tangent approximation cannot always be used. Therefore, we must turn to the methods of differential geometry to represent shapes.
17
~,.
2
•
•
3
1
2
3
•
...~.:;,,,
\
1
•
,
•
/modem human I'
_______Neanderthal "'austr aloptith ecme .
2
3
•
•
I
3
•
•
•
I
\chimpanzee
2
•
~
h  human n  Neanderthal a  australopithecine
1.4.2 Dinosaur Trackways From Mt. Tom, Massachusetts The statistical theory of shape is particularly concerned with the study of random shapes, and shape comparisons in the presence of random variation in shape. Why should a theory of shape incorporate stochastic assumptions? Let us consider two examples in this and the following section. Consider Figure 1.4, which shows the footprints of dinosaurs of the Late Triassic/Early Jurassic period at the Mt. Tom site north of Holyoke, Massachusetts. This data set is described by Ostrom [130]. One of the interesting features of this data set is the presence of multiple tracks that are sufficiently separated to permit the examination of • variation of tracks along the path of a single dinosaur; • variation of tracks between dinosaurs of the same species;
c
o
0.1
a
n
h
• •
•
•
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
FIGURE 1.3. Preshapes of the four skulls plotted on a circle (above), and with a tangent approximation (below). Also marked on the circle are the six preshapes of configurations of equally spaced points for reference purposes.
18
1.
1.4 A Few More Examples
Introduction .'~
19
and, to a certain extent, • variation in tracks between species.
1
.:. ..~" y
. .
••
.. e. .. . / '
./
:,
4
... ..... 0:7 .64 1 e. ••
••
5
N
• _..../8 • •
.
A
t ;\. • • • • \:."I.
•
•• ·1· •
· 9
•
•
__ : .e
/ ~_
C
.
12
•
10~ ~3
,. ~
"~ ::.~~"
'.i
15
.i>:
"
•....
•
•
~
~20 ~_.~
10"",
:;..
19
• •• • ••
...
•
.
21
~
22
 .. ~ 2! · ..:/:26 •
•
Multiple comparisons between species and individuals are possible when footprints can be clearly delineated as belonging to different dinosaurs or different species. For example, Ostrom was struck by the tendency of most of the dinosaur tracks to go in roughly the same direction. He considered the evidence from this site and others for the possible gregarious behavior of dinosaurs. The question of whether dinosaurs had any tendency to congregate in packs or herds is an interesting problem within a much larger issue. Experts have long recognized that.dinosaurs had a combination of reptilian and avian features. In modern animals, gregarious behavior is most commonly found in birds rather than reptiles. So any evidence for such behavior would support a more avian interpretation of dinosaurs. In examining the site, Ostrom found indications of twentyeight trackways made of three distinct types of footprints: large broad footprints identified as made by Eubrontes., intermediate size prints resembling those made by Anchisauripus, and small prints identified as made by Grallator. Each of the twentyeight trackways was assigned an overall direction, and these directions were examined within and between species. The trackway directions were classified into two types: those tracks pointing in a roughly westerly direction ranging through an angle of about 30° and sundry directions far removed from the westerly trackways. The fact that the majority of the trackways point in a westerly direction is suggestive of herding behavior. However, we must be cautious with this conclusion. We cannot automatically conclude that the directionality is due to herding because we do not know about the presence of other external agencies that might have forced the dinosaurs in this direction. A more reliable indicator is any possible relationship between species (as determined by footprint classification) and behavior (as determined by track direction). Ignoring trackway 13, which consists of a single print pointing south and whose identification as Eubrontes is suspect, we can classify the trackways using a 2 x 2 table as follows.
24 25
21:··· / • •
• •
FIGURE 1.4. Dinosaur footprints at the Mt. Tom site near Holyoke, Massachusetts. Footprints can be grouped in partly overlapping trackways corresponding to three species of dinosaurs.
Track Eubrontes Other
West
Other
19
3
14
A simple method for detecting the presence of gregarious behavior from this table is to test for independence between species, listed vertically, and direction, listed horizontally. So the null hypothesis that gregarious behavior is absent can be modeled by the hypothesis of independence of rows and
16
1. Introduction
1.4 A Few More Examples
Michael Lewis of the University of Waterloo. Upon examination, the four skulls are seen to vary particularly in the ratio of the size of the cranium to the size of the jaw. In the human skull this ratio is the largest, while it is smallest for the chimpanzee. The landmarks Xl, X2, and X3 capture some of this variation because the craniumtojaw ratio is proportional to the ratio of the distances from Xl to X2 and from X2 to X3. As Xl, X2, and X3 lie along a line in each picture, we can put some coordinates along each line and consider (Xl, X2, X3) to be a vector in R 3 . The preshape r of such a vector will then be an element of the unit circle 8 1 • Figure 1.3 shows the preshapes of these four configurations of three points plotted on a circle. The reader may be surprised by the small amount of arc length enclosed within the range of the four preshape points in Figure 1.3. This is quite typical of landmarks chosen on biological organisms. Usually, the amount of variation of landmark coordinates between images is small compared to the distances between the landmarks within an image. A small arc of a circle can be approximated by a line segment. So it is tempting to approximate the positions of preshapes on the circle in Figure 1.3 by a similar configuration along a straight line. Such an approximation is called a tangent approximation, and works quite well for many biological data sets. More generally, however, configurations of points on a circle cannot be approximated by a configuration of points along a line without major distortion of the interpoint distances. Similarly, a configuration of points on a shape space such as E~ cannot be approximated by a multivariate configuration in R 2 n  4 without distorting the interpoint Procrusteandistances. So it is fortunate when such a tangent approximation is possible, because it permits the researcher to apply the large collection of multivariate statistical techniques designed for data in Euclidean space. In general, the tangent approximation cannot always be used. Therefore, we must turn to the methods of differential geometry to represent shapes.
2
1
•
•
3
1
3
•
2
•
3
•
•
1
•
1
•
17
3
•
•
/modem human ________Neanderthal \ "australopithecine
3
2
•
•
3
•
chimpanzee
2
•
•
•
h  human n  Neanderthal a  australopithecine
1.4.2 Dinosaur Trackways From Mt. Tom, Massachusetts The statistical theory of shape is particularly concerned with the study of random shapes, and shape comparisons in the presence of random variation in shape. Why should a theory of shape incorporate stochastic assumptions? Let us consider two examples in this and the following section. Consider Figure 1.4, which shows the footprints of dinosaurs of the Late Triassic/Early Jurassic period at the Mt. Tom site north of Holyoke, Massachusetts. This data set is described by Ostrom [130]. One of the interesting features of this data set is the presence of multiple tracks that are sufficiently separated to permit the examination of • variation of tracks along the path of a single dinosaur; • variation of tracks between dinosaurs of the same species;
c
o
0.1
a
n
h
• •
•
•
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
FIGURE 1.3. Preshapes of the four skulls plotted on a circle {above}, and with a tangent approximation (below). Also marked on the circle are the six preshapes of configurations of equally spaced points for reference purposes.
18
1.4 A Few More Examples
1. Introduction
19
and, to a certain extent, ;. ~:~
N
\
..•
..... »>: •
15
19
10"", : ~ •• •••. '.~20 ~_
.......
•
~
U
.c>:
• 22  ...~ .
;
20 •••
';:25
27::::/26 • • • •
FIGURE 1.4. Dinosaur footprints at the Mi. Tom site near Holyoke, Massachusetts. Footprints can be grouped in partly overlapping trackways corresponding to three species of dinosaurs.
• variation in tracks between species.
,
Multiple comparisons between species and individuals are .possibl(3 when footprints can be clearly delineated as belonging to different dinosaurs or different species. For example, Ostrom was struck by the tendency of most of the dinosaur tracks to go in roughly the same direction. He considered the evidence from this site and others for the possible gregarious behavior of dinosaurs. The question of whether dinosaurs had any tendency to congregate in packs or herds is an interesting problem within a much larger issue. Experts have long recognized that dinosaurs had a combination of reptilian and avian features. In modern animals, gregarious behavior is most commonly found in birds rather than reptiles. So any evidence for such behavior would support a more avian interpretation of dinosaurs. In examining the site, Ostrom found indications of twentyeight trackways made of three distinct types of footprints: large broad footprints identified as made by Eubrontes, intermediate size prints resembling those made by Anchisauripus, and small prints identified as made by Grallator. Each of the twentyeight trackways was assigned an overall direction,and these directions were examined within and between species. The trackway directions were classified into two types: those tracks pointing in a roughly westerly direction ranging through an angle of about 30° and sundry directions far removed from the westerly trackways. The fact that the majority of the trackways point in a westerly direction is suggestive of herding behavior. However, we must be cautious with this conclusion. We cannot automatically conclude that the directionality is due to herding because we do not know about the presence of other external agencies that might have forced the dinosaurs in this direction. A more reliable indicator is any possible relationship between species (as determined by footprint classification) and behavior (as determined by track direction). Ignoring trackway 13, which consists ofa single print pointing south and whose identification as Eubrontes is suspect, we can classify the trackways using a 2 x 2 table as follows.
Track
~ West
Eubrontes ~ 19 Other .I=1
Other
3 4
''
A simple method for detecting the presence of gregarious behavior from this table is to test for independence between species, listed vertically, and direction, listed horizontally. So the null hypothesis that gregarious behavior is absent can be modeled by the hypothesis of independence of rows and
20
1. Introduction
1.4 A Few More Examples
columns. A test for independence on this 2 x 2 table is quite significant, and in favor of the hypothesis that there is gregarious behavior. However, we must be cautious in our conclusions because other factors could affect the relationship between track direction and species other than the herding hypothesis. More generally, we might seek to model dinosaur movements across the area so as to make inferences about differences between individuals within species and between species. Quite a large number of footprints of Eubrontes are available. In track 1, for example, the footprints are clearly defined as belonging to a single Eubrontes, and can be interpreted in order as a sequence of successive footprints. Can we use this and similar tracks to model dinosaur motion? We can model a sequence of consecutive footprints as generated by some appropriate random mechanism and then attempt to make inferences by decomposing the geometric configuration of footprints in a trackway into orientation, size, and shape information. We have already performed a rough analysis of the orientations in considering herding behavior. In Chapter 6, we shall consider how size information, available through stride length, can be used to estimate the speed with which the dinosaurs crossed the site. Finally, we shall perform a shape analysis on the trackways and in particular shall investigate how the shape of the triangle formed by three successive footprints is correlated with size variables such as stride length. The unifying approach to such data sets will be to decompose the geometric information into its orientation, size, and shape components, and to consider the variation in these components and their relation to each other.
1.4.3
• v
• N
f
Q
•
Late Bronze Age Post Mold Configurations in England
Consider the configuration of post molds from two Late Bronze Age sites at Aldermaston Wharf and at South Lodge camp in Wiltshire, England. See Figures 1.5 and 1.6. In archeological excavations, clear evidence is often found for the existence of wooden buildings at the site through the configurations of supporting posts of the structure. While these posts are no longer present at the site, the positions of many of them can be determined from the presence of round discolorations of the soil beneath the surface. These discolorations, or post molds, are often found in a roughly regular geometric pattern that indicates the presence of a wall. However, complications can arise in interpreting the post mold evidence. Destructive processes such as erosion can prevent post molds from being detected. Sometimes a building at a particular location was demolished and a succession of other buildings erected at the same place. In these cases, the superimposed post mold patterns can be very difficult to disentangle. From Figures 1.5 and 1.6 we see such problems. It is known that typical buildings of the time were circular structures called roundhouses. Neighboring posts were usually 1.6 to 2.2 meters apart, and possibly up to three meters apart. In Figure 1.6, the
21
o
a
000
o
o
•
•
On""
10m
FIGURE 1.5. Post mold configuration at Aldermaston Wharf showing links between neighboring post molds. Later features are marked as shaded reqions. Irregular unshaded regions are pits at the site. This figure is adapted from [32} by kind permission of The Museum Applied Science Center at the University of Pennsylvania.
22
1.
Introduction
1.4
( ,
o .>. a
Om
o
une'COVOlr7f}ed
o o
i
i
I
,
.,/
I'
/
10m
FIGURE 1.6. Post mold configuration at South Lodge Camp showing links between neighboring post molds. A large, highly regular circular configuration of post molds can be seen on the east side of the site. A smaller circle adjacent to it is also visible~ This figure is adapted from [32} by kind permission of The Museum Applied Science Center at the University of Pennsylvania.
.~
.
A Few More Examples
23
post molds whose interpoint distances are less than three meters have been linked by a line segment. Four main clusters of points, labeled A, B, C, and D can be seen. Strong visual evidence for the existence of a roundhouse can be seen in cluster D of the outline plan of South Lodge Camp. The clearly circular arrangement of posts would be difficult to explain as a coincidence from a purely random mechanism. On the other hand, the evidence from cluster C is more ambiguous. Here there is also some indication of a roundhouse. However, in this case it is more difficult to determine whether the circular pattern is too regular to arise simply by chance. Finally, in cluster .A there is a very slight indication of a roundhouse. But here we would have to admit that any evidence of a circle could quite possibly be coincidental. There is no clear confirmation that a circular building was present here, although there is a suggestion of circularity in the positions of the post molds. At Aldermaston Wharf, the evidence for circular buildings is provided by the positions of post molds clustered visually as Structure I and Structure II in Figure 1.5. Of these two, Structure II is the better formed, and has six post molds that can be placed on a rough circle. Structure I looks very irregular. Again, there are six post molds that can be interpreted as circular. Neither structure is as compelling as Cluster D from South Lodge Camp. How should we assess the patterns at these two sites, and how can we determine whether such configurations are likely by chance in a random scattering? A method for fitting circles that is particularly amenable to analysis of this kind has been provided by Cogbill [44]. He proposed that circular configurations of posts can be detected by running an annulus across the window in which the posts are plotted. If the inner and outer radii of the annulus are close, the thin annulus will cover few points in any given position. However, by chance, at certain positions a larger number of points will be covered. Such configurations of. posts can be examined for the possibility that they form the circular boundary of a roundhouse. For example, the six points of Structure II at Aldermaston Wharf can be completely contained in an annulus whose inner radius is 3.66 meters and whose outer radius is 3.95 meters. Is such a fit likely by chance? We could define chance configurations as those arising in a random uniform scattering of equally many points over a similar region. In such a scattering, what is the expected number of circles that will be found of six points covered by an annulus of inner and outer.radii3.66 and 3.95 meters respectively? Early work by Mack [111] provides a powerful tool for answering this question. In Chapter 6, we shall see that we would expect to discover a circular arrangement of this tolerance simply by chance if the posts were randomly scattered across the region of excavation. Such a calculation casts doubt upon the strength of the archeological interpretation at Aldermaston. A similar analysis of Cluster D at South Lodge Camp is more reassuring for archeological interpretation. In this case, a set of eight points can be fit with
20
1.
Introduction
columns. A test for independence on this 2 x 2 table is quite significant, and in favor of the hypothesis that there is gregarious behavior. However, we must be cautious in our conclusions because other factors could affect the relationship between track direction and species other than the herding hypothesis. More generally, we might seek to model dinosaur movements across the area so as to make inferences about differences between individuals within species and between species. Quite a large number of footprints of Eubronies are available. In track 1, for example, the footprints are clearly defined as belonging to a single Eubrotites, and can be interpreted in order as a sequence of successive footprints. Can we use this and similar tracks to model dinosaur motion? 'vVe can model a sequence of consecutive footprints as generated by some appropriate random mechanism and then attempt to make inferences by decomposing the geometric configuration of footprints in a trackway into orientation, size, and shape information. We have already performed a rough analysis of the orientations in considering herding behavior. In Chapter 6, we shall consider how size information, available through stride length, can be used to estimate the speed with which the dinosaurs crossed the site. Finally, we shall perform a shape analysis on the trackways and in particular shall investigate how the shape of the triangle formed by three successive footprints is correlated with size variables such as stride length. The unifying approach to such data sets will be to decompose the geometric information into its orientation, size, and shape components, and to consider the variation in these components and their relation to each other.
1.4.3
1.4 A Few More Examples
.v
0
~
~. o
o
o
tao 0
0
o
o
0
•
~
:.
0
•
c
0""
•
~.,
N
D
• \)0
0
O· /°6
o
0
0
Late Bronze Age Post Mold Configurations in England
Consider the configuration of post molds from two Late Bronze Age sites at Aldermaston Wharf and at South Lodge camp in Wiltshire, England. See Figures 1.5 and 1.6. In archeological excavations, cl~ar evidence is often found for the existence of wooden buildings at the site through the configurations of supporting posts of the structure. While these posts are no longer present at the site, the positions of many of them can be determined from the presence of round discolorations of the soil beneath the surface. These discolorations, or post molds, are often found in a roughly regular geometric pattern that indicates the presence of a wall. However, complications can arise in interpreting the post mold evidence. Destructive processes such as erosion can prevent post molds from being detected. Sometimes a building at a particular location was demolished and a succession of other buildings erected at the same place. In these cases, the superimposed post mold patterns can be very difficult to disentangle. From Figures 1.5 and 1.6 we see such problems. It is known that typical buildings of the time were circular structures called roundhouses. Neighboring posts were usually 1.6 to 2.2 meters apart, and possibly up to three meters apart. In Figure 1.6, the
• 0
o ... ~
21
D
000
0
t>.
~ t .
10m
FIGURE 1.5. Post mold configuration at Aldermaston Wharf showing links between neighboring post molds. Later features are marked as shaded regions. Irregular unshaded regions are pits at the site. This figure is adapted from [32} by kind permission of The Museum Applied Science Center at the University of Pennsylvania.
22
1.
Introduction
1.4
• •
o a
Om
. : •.,une'COVO'f11)ed i
i I
;
/
I'
o
o o
/
10m !] :
FIGURE 1.6. Post mold configuration at South Lodge Camp showing links between neighboring post molds. A large, highly regular circular configuration of post molds can be seen on the east side of the site. A smaller circle adjacent to it is also visible. This figure is adapted from [32] by kind permission of The Museum Applied Science Center at the University of Pennsylvania;
'
..
;j
A Few More Examples
23
post molds whose interpoint distances are less than three meters have been linkedbya line segment. Four main clusters of points, labeled A, B, C, and D can be seen. Strong visual evidence for the existence of a roundhouse can be seen in cluster D of the outline plan of South Lodge Camp. The clearly circular arrangement of posts would be difficult to explain as a coincidence from a purely random mechanism. On the other hand, the evidence from cluster C is more ambiguous. Here there is also some indication of a roundhouse. However, in this case it is more difficult to determine whether the circular pattern is too regular to arise simply by chance. Finally, in cluster A there is a very slight indication of a roundhouse. But here we would have to admit that any evidence of a circle could quite possibly be coincidental. There is no clear confirmation that a circular building was present here, although there is a suggestion of circularity in the positions of the post molds. At Aldermaston Wharf, the evidence for circular buildings is provided by the positions of post molds clustered visually as Structure I and Structure II in Figure 1.5. Of these two, Structure II is the better formed, and has six post molds that can be placed on a rough circle. Structure I looks very irregular. Again, there are six post molds that can be interpreted as circular. Neither structure is as compelling as Cluster D from South Lodge Camp. How should we assess thepatterns at these two sites, and how can we determine whether such configurations are likely by chance in a random scattering? A method for fitting circles that is particularly amenable to analysis of this kind has been provided by Cogbill [44]. He proposed that circular configurations of posts can be detected by running an annulus across the window in which the posts are plotted. If the inner and outer radii of the annulus are close, the thin annulus will cover few points in any given position. However, by chance, at certain positions a larger number of points will be covered. Such configurations of posts can be examined for the possibility that they form the circular boundary of a roundhouse. For example, the six points of Structure II at Aldermaston Wharf can be completely contained in an annulus whose inner radius is 3.66 meters and whose outer radius is 3.95 meters. Is such a fit likely by chance? We could define chance configurations as those arising in a random uniform scattering of equally many points over a similar region. In such a scattering, what is the expected number of circles that will be found of six points covered by an annulus of inner and outer radii 3.66 and 3.95 meters respectively? Early work by Mack [111] provides a powerful tool for answering this question. In Chapter 6, we shall see that we would expect to discover a circular arrangement of this tolerance simply by chance if the posts were randomly scattered across the region of excavation. Such a calculation casts doubt upon the strength of the archeological interpretation at Aldermaston. A similar analysis of Cluster D at South Lodge Camp is more reassuring for archeological interpretation. In this case, a set of eight points can be fit with
24
1. Introduction
1.5 The Problem of Homology
25
an annulus with inner radius 3.95 meters and outer radius 4.21 meters. As we shall see, we expect such circular arrangements in a comparable random scattering less than one time in six. Even this looks rather high in view of the precision of the circle of points in Cluster D. However, the circular fit does not take into account the even spacing of posts, which is also unlikely in a random scattering.
1.5
The Problem of Homology
In the biological sciences, sites or landmarks on different organisms are said to be homologous if they share a common structure and evolutionary origin. For example, the eyes of a chimpanzee arc homologous to the eyes of a human despite the shape differences between the head of a chimpanzee and the head of a human. More generally, outside the biological sciences, sites on different bodies or images are said to be homologous if they naturally correspond due to a common structure. We considered an example of this in Section 1.3, where we chose four landmarks on each of three Iron Ages brooches so that correspondingly labeled landmarks were homologous between images. Homologous landmarks are not always obvious, and may depend upon insight or expert opinion for their construction. As an illustration of the problems associated with constructing satisfactory homologies between images, let us consider the work of Thompson [172], who devised a method for examining shape differences between biological organisms called the method of coordinates. The reader can find an example of Thompson's method by looking at Figure 1.7. In this figure, we see four lateral views of the skulls that. we considered in Section 1.4.1 and Figure 1.2. Thompson proposed the placement of a rectangular grid over one of the images, say the modern human skull at the top. Now, each of the intersection points of the grid corresponds to a feature of some kind in the skull. (The detection of such features requires more detailed information than is available in Figure 1.7.) Suppose that for each feature at every intersection point in the top grid we are able to find the corresponding (homologous) feature in the other skulls. A horizontal or vertical line of the Cartesian grid on the top image is mapped to a curvilinear line in each of the other images by connecting sites in the other images that are homologous to sites on the same horizontal or vertical line of the top image. The resulting coordinate system superimposed on the second image is typically curvilinear in nature. The degree to which the curvilinear coordinate systems depart from a Cartesian frame is a measure of the shape differences between the images. By looking at the curvilinear coordinate systems of Figure 1.7, we can make some detailed observations about the shape variation among the four skulls. In particular, by looking at the upper left and lower right corners,
=1Q
;;a;:
.u:
~
1
t I
I
I
FIGURE 1.7. Side view of skulls. From top to bottom: modern human, Neanderthal, australopithecine, chimpanzee. To the right of each skull is a coordinate grid determined with Thompson's method of coordinates, with the modern human skull as the base image. Reproduced from Figure 3.53 of {l3i] by kind permission of Hong Kong University Press.
1.7 Problems
1. Introduction we can see what was observed in Section 1.4.1, namely that an important source of variation is to be found in the change in the relative sizes of jaw and cranium. With far more coordinates available for comparison, we are able to make a much more detailed examination of these differences. However, Thompson's method of coordinates has several problems. The first of these is the problem of how to draw a smooth line appropriately through a set of points. This is essentially an interpolation, or fitting problem. The second problem is that mentioned above, namely of finding a correspondence, or homology, between landmarks on different images. A final problem is to decide how to summarize the information available about the differences in shapes among the images from such complicated grids of curvilinear coordinates. We will consider these problems again in Chapter 3.
1.6 Notes The theory of shape owes much to D' Arcy Thompson [172] for its inspiration. His work has long been regarded as a model for the fusion of scientific, mathematical, and literary skills. Although his analyses of biological growth and form are now dated, his exposition of the theory of biological shape is unparalleled for its clarity. The reader who has not encountered his work is strongly encouraged to do so. For a comprehensive discussion of the theory and methods of morphometries, the reader is referred to [139]. Brief surveys of allometric methods are to be found in [81] and [125]. A variety of applications is readily available in the literature, including [7], [10], [13], [24], [45], [63], [85], [107], and [126], to name afew. The mathematical theory of shape that has been introduced in this chapter can befound in Kendall [90]. This paper was seminal for the development of .this particular school of shape theory, which can be called the Kendall school or perhaps the Procrustean school of shape analysis. Much of the; material in the following chapters relies on the Kendall school of shape and takes advantage of its comprehensive methodology for the analysis of finite point sets in arbitrary dimensions. In particular, the definition and metric of I;;, the space of shapes of n points in p dimensions, is due to Kendall. For extensions of Kendall's work to more general multivariate normal models, see [53]. The Bookstein school of shape analysis uses a different geometric structure on shape spaces that will be discussed in Chapter 3. As mentioned earlier, we have used the word landmark in a more general sense than Bookstein as a point chosen from a body that helps summarize its geometric features. Bookstein has recommended the use of landmarks for the analysis of biological features and constrains the choice of landmarks to
27
prominent features of the organism or biological structure. For the analysis of more general shapes outside the biological sciences, the choice of natural sites for landmarks remains a desirable goal, but is very restrictive for shape description. Therefore, we choose a generalized interpretation of landmark data. A synthesis of the Kendall, or Procrustean, school of shape with the use of landmarks can be found in the survey paper of Goodall [66]. An alternative approach to the selection of landmarks can be found in [60].
~ 
1.7 Problems 1. A researcher proposes to define the shape of a triangle as a vector (al,a2,a3) of three internal angles. Discuss the advantages and disadvantages of encoding shape information in this way. 2. Two triangles are congruent if their corresponding sides are of equal length. A researcher proposes to encode the size and shape information about a triangle as a vector (d 12, d 13, d 23) E R 3 where djk is the length of the side joining the jth and kth vertices. A size variable W{d 12, d13' d23 ) is a nonnegative function that is homogeneous, in the sense that (1.24) for all t 2: O. Give two distinct examples of size variables and show how shape coordinates for a triangle can be constructed by standardizing the djk with respect to size. 3. The next two problems involve the concept of a random shape statistic. In this problem, the shape statistic in question is the maximum internal angle.of a random triangle. In the next problem, the statistic is an indicator of the event that a random quadrilateral is convex. The reader who is not familiar with the probability theory used in these questions can safely pass over these problems until we return to probability theory in Chapter 4. Three random planar points are independent, with a common absolutely continuous distribution. Let M be the maximum internal angle of the triangle whose vertices are the three points. Show that (1.25) (Hint: consider six such random points.) 4. Four random planar points are independent, with a common absolutely continuous distribution. Show that with probability greater than or equal
24
1.5 The Problem of Homology
1. Introduction
25
an annulus with inner radius 3.95 meters and outer radius 4.21 meters. As we shall see, we expect such circular arrangements in a comparable random scattering less than one time in six. Even this looks rather high in view of the precision of the circle of points in Cluster D. However, the circular fit does not take into account the even spacing of posts, which is also unlikely in a random scattering.
1.5
The Problem of Homology
In the biological sciences, sites or landmarks on different organisms are said to be homologous if they share a common structure and evolutionary origin. For example, the eyes of a chimpanzee are homologous to the eyes of a human despite the shape differences between the head of a chimpanzee and the head of a human. More generally, outside the biological sciences, sites on different bodies or images are said to be homologous if they naturally correspond due to a common structure. We considered an example of this in Section 1.3, where we chose four landmarks on each of three Iron Ages brooches so that correspondingly labeled landmarks were homologous between images. Homologous landmarks are not always obvious, and may depend upon insight or expert opinion for their construction. As an illustration of the problems associated with constructing satisfactory homologies between images, let us consider the work of Thompson [172J, who devised a method for examining shape differences between biological organisms called the method of coordinates. The reader can find an example of Thompson's method by looking at Figure 1.7. In this figure, we see four lateral views of the skulls that we considered in Section 1.4.1 and Figure 1.2. Thompson proposed the placement of a rectangular grid over one of the images, say the modern human skull at the top. Now, each of the intersection points of the grid corresponds to a feature of some kind in the skull. (The detection of such features requires more detailed information than is available in Figure 1.7.) Suppose that for each feature at every intersection point in the top grid we are able to find the corresponding (homologous) feature in the other skulls. A horizontal or vertical line of the Cartesian grid on the top image is mapped to a curvilinear line in each of the other images by connecting sites in the other images that are homologous to sites on the same horizontal or vertical line of the top image. The resulting coordinate system superimposed on the second image is typically curvilinear in nature. The degree to which the curvilinear coordinate systems depart from a Cartesian frame is a measure of the shape differences between the images. By looking at the curvilinear coordinate systems of Figure 1.7, we can make some detailed observations about the shape variation among the four skulls. In particular, by looking at the upper left and lower right corners,
'L

\~
\
~~ ~ .~
l
o.
J L...
J
I
I
FIGURE 1.7. Side view of skulls. Prom top to bottom: modern human, Neanderthal, australopithecine, chimpanzee. To the right of each skull is a coordinate grid determined with Thompson's method of coordinates, with the modern human skull as the base image. Reproduced from Figure 3.53 of [131/ by kind permission of Hong Kong University Press.
26
1.
we can see what was observed in Section 1.4.1, namely that an important source of variation is to be found in the change in the relative sizes of jaw and cranium. With far more coordinates available for comparison, we are able to make a much more detailed examination of these differences. However, Thompson's method of coordinates has several problems. The first of these is the problem of how to draw a smooth line appropriately through a set of points. This is essentially an interpolation, or fitting problem. The second problem is that mentioned above, namely of finding a correspondence, or homology, between landmarks on different images. A final problem is to decide how to summarize the information available about the differences in shapes among the images from such complicated grids of curvilinear coordinates. We will consider these problems again in Chapter 3.
1.6
1.7 Problems
Introduction
Notes
The theory of shape owes much to D' Arcy Thompson [172] for its inspiration. His work has long been regarded as a model for the fusion of scientific, mathematical, and literary skills. Although his analyses of biological growth and form are now dated, his exposition of the theory of biological shape is unparalleled for its clarity. The reader who has not encountered his work is strongly encouraged to do so. For a comprehensive discussion of the theory and methods of morphometrics, the reader is referred to [139]. Brief surveys of allometric methods are to be found in [81] and [125]. A variety of applications is readily available in the literature, including [7], [10], [13], [24]' [45], [63], [85], [107], and [126], to name a few. The mathematical theory of shape that has been introduced in this chapter can be found in Kendall [90]. This paper was seminal for the development of this particular school of shape theory, which can be called the Kendall school or perhaps the Procrustean school of shape analysis. Much of the material in the following chapters relies on the Kendall school of shape and takes advantage of its comprehensive methodology for the analysis of finite point sets in arbitrary dimensions. In particular, the definition and metric of I;;, the space of shapes of n points in p dimensions, is due to Kendall. For extensions of Kendall's work to more general multivariate normal models, see [53]. The Bookstein school of shape analysis uses a different geometric structure on shape spaces that will be discussed in Chapter 3. As mentioned earlier, we have used the word landmark in a more general sense than Bookstein as a point chosen from a body that helps summarize its geometric features. Bookstein has recommended the use of landmarks for the analysis of biological features and constrains the choice of landmarks to
27
prominent features of the organism or biological structure. For the analysis of more general shapes outside the biological sciences, the choice of natural sites for landmarks remains a desirable goal, but is very restrictive for shape description. Therefore, we choose a generalized interpretation of landmark data. A synthesis of the Kendall, or Procrustean, school ofshape with the use of landmarks can be found in the survey paper of Goodall [66]. An alternative approach to the selection of landmarks can be found in [60].
1.7
Problems
,;;,
1. A researcher proposes to define the shape of a triangle as a vector (aI, a2, (3) of three internal angles. Discuss the advantages and disadvantages of encoding shape information in this way.
'j
.J;
2. Two triangles are congruent if their corresponding sides are of equal length. A researcher proposes to encode the size and shape information about a triangle as a vector (d I 2 , d l 3 , d 23 ) E R 3 where djk is the length of the side joining the jth and kth vertices. A size variable W(d 12 , d 13 , d23 ) is a nonnegative function that is homogeneous, in the sense that (1.24) for all t:2: O. Give two distinct examples of size variables and show how shape coordinates for a triangle can be constructed by standardizing the djk with respect to size. 3. The next two problems involve the concept of a random shape statistic. In this problem, the shape statistic in question is the maximum jntemal angle of a random triangle. In the next problem, the statistic is an indicator of the event that a random quadrilateral is convex. The reader who is not familiar with the probability theory used in these questions can safely pass over these problems until we return to probability theory in Chapter 4. Three random planar points are independent, with a common absolutely continuous distribution. Let M be the maximum internal angle of the triangle whose vertices are the three points. Show that (1.25) (Hint: consider six such random points.) 4. Four random planar points are independent, with a common absolutely continuous distribution. Show that with probability greater than or equal
28
1.
Introduction
to 1/5 one of the four points will lie in the triangle formed by the other three.
2
5. In formula (1.21) we encountered the Procrustean metric. A metric d(x, y) between points x, y of a set is a nonnegative real valued function satisfying (i) d(x, y) = a if and only if x = Yi (ii) d(x, y) = dey, x) for all x and Yi (iii) d(x, z) ::; d(x, y) + dey, z) for all x, y, and z, Show that the Procrustean metric d defined in Section 1.3 satisfies these properties on the set L:2'.
Background Concepts and Definitions
2.1 ~
Transformations on Euclidean Space
In this section, we shall begin with some preliminary definitions relevant to shape analysis.
2.1.1
Properties of Sets
Let RP be the usual pdimensional Euclidean space. A subset A c RP is said to be open if for every x E A, there is some to > a such that yEA whenever Ilx  vll < e. A subset A is said to be closed if its complement Ac in RP is open. By the interior AO of any subset A we mean the largest open subset of A, possibly the empty set. The interior of A is found as the union of all open subsets of A. A subset A c RP is said to be convex if for every x, YEA, the line segment with endpoints at x and y lies entirely in A. The convex hull of any given A c RP is the smallest convex set that contains A. The convex hull of A is found as the intersection of all convex sets that contain the set A.
2.1.2
Affine Transformations
Let A = (Aj l,) be a q x p matrix. By a linear transformation from RP to Rq we shall mean a mapping of the form x > Ax, where x is a p x 1 column vector. Linear transformations are special cases of affine
30
2.1 Transformations on Euclidean Space
2. Background Concepts and Definitions
transformations, which have the general form x + Ax + a, where a is any p x 1 column vector. Suppose that Xl, .•• , X p+! are P + 1 points in RP. These points form the vertices of a psimplex in RP, which can be defined as the convex hull of these points. Suppose Xl, ... ,xp+! and Yl, .•. , Yp+! are the vertices of two psimplexes with positive pdimensional volume. Then there exists a unique affine transformation RP + RP of the form x + Ax + a such that Yj=Axj+a for all j=1,2, ...,p+1.
2.1.3
is a column vector of p complex values and A is a p xp unitary matrix, is said to be a unitary transformation of CP. For any unitary matrix A the determinant det(A) is a complex number with modulus one. We say that A is a special unitary matrix provided det(A) = l. Just as the complex plane C can be identified with R2, so the unitary transformations of CP can be identified with particular orthogonal transformations of R 2P . The 2p x 2p matrix of real values corresponding to the unitary matrix A is found by replacing each complex entry A j k by the 2 x 2 block of real values
Orthogonal Transformations
A p x p matrix A = (Ajk) is said to be orthogonal if AT = A\ where AT and AI denote the transpose and inverse matrices of A respectively. Equivalently, we can say that AT A = I, where I is the p x p identity matrix. By an orthogonal transformation from RP to itself we shall mean a linear transformation x + Ax corresponding to multiplication of a pdimensional column vector on the left by a p x p orthogonal matrix. For any orthogonal matrix A the determinant det(A) = ±1. Those orthogonal matrices with det(A) = 1 are said to be special orthogonal matrices, and their corresponding transformations of RP are said to be special orthogonal transformations. Special orthogonal transformations can be regarded as generalizations into higher dimensions of the families of rotations about the origin in dimensions two and three. An example of an orthogonal transformation that is not a special orthogonal transformation is the reflection (2.1)
of RP through the hyperplane Xl = O. Henceforth, we shall let O(p) and SO(p) denote the classes of orthogonal and special orthogonal transformations on RP respectively.
2.1.4
31
Unitary Transformations
We now describe an analog to the class of orthogonal transformations on RP. Let C be the complex plane, and CP the :space. of pvectors whose entries are elements of C. Linear transformations from RP to RP can be represented as x + Ax, where A is a p x p matrix of real entries. The complex analogs of these transformations are also of the form x+ Ax, with the real entries of the column vector x and the matrix A replaced by complex values. These are linear transformations from CP to CPo Suppose A = (Aj k ) is a p x p matrix of complex values. Let A* be the p x p matrix whose (j, k)th entry is the complex conjugate of the (k,j)th entry of A. Then A is said to be a unitary matrix if A* A = I, where I is the p x p identity matrix. A linear transformation x + Ax, where x
(2.2) where ~(z) and SS(z) are the real and imaginary parts of the complex number z respectively. Thus every unitary transformation of CP can be regarded as an orthogonal transformation of R 2P. Under this identification, the determinant ofthe 2px2p orthogonal matrix will be the modulus of the determinant of its p x p unitary counterpart. While every unitary matrix or transformation can be identified with an orthogonal transformation, the converse is not true. This follows easily from the previous observation that the determinant of its 2p x 2p orthogonal counterpart equals one, being the modulus of a complex number on the unit circle of C. Thus reflections of R 2 p , and many other orthogonal transformations, cannot be represented as unitary transformations. Henceforth, we shall let U(p) and SU(p) denote, respectively, the classes of unitary and special unitary transformations on CP.
2.1.5
Singular Value Decompositions
Let A be a matrix of dimension q x p that has rank r. Then AAT (or equivalently AT A) has r nonzero eigenvalues. It is easy to check that the eigenvalues of AAT are nonnegative. Therefore we can write the eigenvalues as At, A~, ..., A;. We define the matrix r = (r j k ) to be a q xp matrix for which r j j = IAjl for j = l,2, ... ,r and whose other elements are zero. Then A can be written as
(2.3) where Wand W' are orthogonal matrices of dimension q x q and p x p respectively. This decomposition is called a singular value decomposition of A. The eigenvalues IA1I, ...,IArl are called the singular values of the matrix A. Note that the singular value decomposition of A is not unique, although the set of singular values of A is uniquely determined.
28
1.
Introduction
to 1/5 one of the four points will lie in the triangle formed by the other three. 5. In formula (1.21) we encountered the Procrustean metric. A metric d(x,y) between points x,y of a set is a nonnegative real valued function satisfying (i) d(x, y) = 0 if and only if x = y; (ii) d(x, y) = d(y, x) for all x and y; (iii) d(x, z) ::s: d(x, y) + d(y, z) for all x, y, and z. Show that the Procrustean metric d defined in Section 1.3 satisfies these properties on the set E2'.
2 Background Concepts and Defini tions
2.1
Transformations on Euclidean Space
In this section, we shall begin with some preliminary definitions relevant to shape analysis.
2.1.1
Properties of Sets
Let RP be the usual pdimensional Euclidean space. A subset A c RP is said to be open if for every x E A, there is some 10 > 0 such that yEA whenever Ilx  yll < 10. A subset A is said to be closed if its complement A C in RP is open. By the interior A 0 of any subset A we mean the largest open subset of A, possibly the empty set. The interior of A is found as the union of all open subsets of A. A subset A c RP is said to be convex if for every x, yEA, the line segment with endpoints at x and y lies entirely in A. The convex hull of any given A c RP is the smallest convex set that contains A. The convex hull of A is found as the intersection of all convex sets that contain the set A.
2.1.2
Affine Transformations
Let A = (A j k ) be a q x p matrix. By a linear transformation from RP to Rq we shall mean a mapping of the form x + Ax, where x is a p x 1 column vector. Linear transformations are special cases of affine
30
2.1 Transformations on Euclidean Space
2. Background Concepts and Definitions
transformations, which have the general form x > Ax + a, where a is any p x 1 column vector. Suppose that Xl, ... , x p +! are P + 1 points in RP. These points form the vertices of a psimplex in RP, which can be defined as the convex hull of these points. Suppose z i , ... ,Xp+l and Yl, ... ,Yp+!· are the vertices of two psimplexes with positive pdimensional volume. Then there exists a unique affine transformation RP > RP of the form X > Ax + a such that Yj = AXj + a for all j = 1,2, ...,p + 1.
2.1.3
.
is a column vector of p complex values and A is a p x p unitary matrix, is said to be a unitary transformation of CP. For any unitary matrix A the determinant det(A) is a complex number with modulus one. We say that A is a special unitary matrix provided det(A) = 1. J list as the complex plane C can be identified with R 2 , so the unitary transformations of CP can be identified with particular orthogonal transformations of R 2P . The 2p x 2p matrix of real values corresponding to the unitary matrix A is found by replacing each complex entry Ajk by the 2 x 2 block of real values
Orthogonal Transformations
A p x p matrix A = (Ajk) is said to be orthogonal if AT = A\ where AT and AI denote the transpose and inverse matrices of A respectively. Equivalently, we can say that ATA = I, where I is the p x p identity matrix. By an orthogonal transformation from RP to itself we shall mean a linear transformation x > Ax corresponding to multiplication of a pdimensional column vector on the left by a p x p orthogonal matrix. For any orthogonal matrix A the determinant det(A) = ±1. Those orthogonal matrices with det(A) = 1 are said to be special orthogonal matrices, and their corresponding transformations of RP are said to be special orthogonal transformations. Special orthogonal transformations can be regarded as generalizations into higher dimensions of the families of rotations about the origin in dimensions two and three. An example ofan orthogonal transformation that is not a special orthogonal transformation is the reflection (2.1) of RP through the hyperplane Xl = O. Henceforth, we shall let O(p) and SO(p) denote the classes of orthogonal and special orthogonal transformations on RP respectively.
2.1..4
.~
31
Unitary Transformations
We now describe an analog to the dass of orthogonal transformations on RJ'. Let C be the complex plane, and CP the space of pvectors whose entries are elements of C. Linear transformations from RP to RP can be represented as X > Ax, where A is a px p matrix of real entries. The complex analogs of these transformations are also of the form x > Ax, with the real entries of the column vector x and the matrix A replaced by complex values. These are linear transformations from CP to CP. Suppose A = (Aj k ) is a p x p matrix of complex values. Let A* be the p x p matrix whose (j, k)th entry is the complex conjugate of the (k,j)th entry of A. Then A is said to be a unitary matrix if A* A = 1, where I is the p x p identity matrix. A linear transformation x > Ax, where x
(2.2) where ~(z) and SJ'(z) are the real and imaginary parts of the complex number z respectively. Thus every unitary transformation of CP can be regarded as an orthogonal transformation of R 2P . Under this identification, the determinant of the 2p x 2p orthogonal matrix will be the modulus of the determinant of its p x p unitary counterpart. While every unitary matrix or transformation can be identified with an orthogonal transformation, the converse is not true. This follows easily from the previous observation that the determinant of its 2p x 2p orthogonal counterpart equals one, being the modulus of a complex number on the unit circle of C. Thus reflections of R 2p , and many other orthogonal transformations, cannot be represented as unitary transformations. Henceforth, we shall let U(p) and SU(p) denote, respectively, the classes of unitary and special unitary transformations on CP.
2.1.5 Singular Value Decompositions Let A be a matrix of dimension q x p that has rank r, Then AAT (or equivalently ATA) has r nonzero eigenvalues. It is easy to check that the eigenvalues of AAT are nonnegative. Therefore we can write the eigenvalues as A~, A~, .,., A~. We define the matrix r = (r j k ) to be a q xp matrix for which r j j = IAj I for j = 1,2, ... , r and whose other elements are zero. Then A can be written as
A = W r Wi
(2.3)
where Wand Wi are orthogonal matrices of dimension q x q and p x p respectively. This decomposition is called a singular value decomposition of A. The eigenvalues IAll, ..., jArl are called the singular values of the matrix A. Note that the singular value decomposition of A is not unique, although the set of singular values of A is uniquely determined.
32
2. Background C~ncepts and Definitions
2.1 Transformations on Euclidean Space
A case that will be of particular interest to us occurs when p = q and A is of full rank. In this case, r is a square diagonal matrix whose diagonal elements are the singular values. The singular value decomposition allows us to represent a matrix in diagonal form, with i[1 and i[11 serving to provide a change of coordinate systems for the purpose. The singular value decomposition has an important geometric interpretation that will be of use in the next chapter. Suppose x is a 2 x 1 column vector and that A is a 2 x 2 matrix of full rank. Under the linear transformation x > Ax the unit circle in the plane. R 2 is mapped to an ellipse. The lengths of the semimajor and semi minor axes of this ellipse are seen from equation (2.3) to be the singular values of A. This geometric interpretation generalizes into higher dimensions. A p x p matrix A of full rank will have p singular values. If x is a p x 1 column vector, then x , Ax will map the unit sphere in RP to an ellipsoid with p principal axes. The singular values of A can be seen to be one half the lengths of the principal axes of the ellipsoid.
2.1.6
Inner Products
The inner product between two elements, x (Yl, .", Yp) of RP is defined as
(Xl, ..., xp) and Y
P,
< X,Y > = LXiYj
(2.4)
j=l
Its complex counterpart for CP is called the Hermitian inner product. We encountered the Hermitian inner product in Chapter 1 when defining the distance between two shapes in formula (1.21), We define the Hermitian inner product between two vectors X= (Xl""'X p ) and Y = (Yl,""YP) of complex coordinates to be p
«x,y» = LXjyj
(2.5)
j=l
where Y~ is the complex conjugate of Yj E C. Under the identification of C with JR 2 the inner product on R 2p can be defined from the Hermitian inner product on CP by noting that = lR« ". ». Orthogonal transformations can be characterized as linear transformations that preserve inner products. Thus if A = (Aj k ) is an orthogonal matrix, then representing x, Y E RP as column vectors, we have
< Ax, Ay> =
(2.6)
for all x and Y in RP. Similarly, Hermitian inner products on CP are preserved under unitary transformations.
33
2.1.7 Groups of Transformations The classes O(p), SO(p), and their complex analogs U(p) and SU(p) are classes of transformations of a space to itself. We now summarize some definitions and properties of groups, of which these classes are examples. Let h l and h 2 be any two transformations from RP to RP. By the composition of b, and h 2 we shall mean the function h 2 ° h 1 from RP to RP defined by
(2.7)
Suppose h is a 11 function that maps RP onto itself. We shall let h 1 denote the inverse function, where hl(y) = x whenever h(x) = y. There is nothing special about RP in these definitions, as RP can be replaced by CP or any other set. Definition 2.1.1. A nonempty collection H = {h} of 11 transformations on a set is said to be a group provided that it is closed under composition and inversion of transformations. In order for a nonempty collection H of transformations on a set to be a group, it is necessary and sufficient that for any h l , h 2 in H the transformation h 1 0 hi 1 be in H. Setting h l = h 2 we see that the identity transformation e is always an element of H. Definition 2.1.2. Two transformations h 1 and h 2 are said to commute when h 2oh 1 = h 1oh2 • We say that a group H is commutative or Abelian provided that any two elements of H commute. By the center of a group H we mean the set of elements of H that commute with every other element of H. Obviously, a group H is commutative if and only if the center of H is H itself. The class of orthogonal matrices is closed under matrix multiplication as well as matrix inversion. Similarly, the class O(p) of orthogonal transformations is closed under function composition and function inversion. Thus the class of orthogonal transformations is a group, that is commutative only for the cases where p = 1,2. The class SO(p) of special orthogonal transformations is a subgroup of the group of orthogonal transformations. That is, it is a subset of O(p) that is a group in its own right, being also closed under composition and inversion. When p = 1 this subgroup is the trivial group consisting of the identity transformation alone. Similar results hold true for the class of unitary transformations of CPo The class U(p) is also a group, containing the subgroup SU(p).
34
2.1.8
2.
Definition 2.1.3. A 11 correspondence h: M Jo N between metric spaces is said to be an isometry if d(x, y) = d[h(x), h(y)] for all x, y E M. Two metric spaces are said to be isometric if there is an isometry mapping one to the other. When M and N are isometric, we shall write M ~ N. In particular, the class of all isometries fromM to itself shall be denoted Iso(M). It is immediate that the identity transformation on M is an isometry, and it can be checked that the transformations of Iso(M) form a group. On RP, for example, the class of Euclidean motions Euc(p) forms a subgroup of Iso(RP). This subgroup is a proper subgroup, because the Eucidean motions of RP do not include reflections through a (p  1)dimensional hyperplane. We may also speak of a linear isometry between vector spaces. Definition 2.1.4. A linear transformation of full rank between two vector spaces is said to be a linear isometry if it preserves the lengths of vectors. Clearly, an orthogonal rotation of RP is an example of a linear isometry from RP to itself.
Similarity Transformations and the Shape of Sets
Let (Xl, ..., x p ) be an element of RP. A transformation (Xl, ..., x p ) Jo (AXI, ... , AXp ) , where A> 0, is said to be an isotropic rescaling or simply a rescaling of RP. By a shapepreserving transformation ora similarity trans
35
formation of RP, we shall mean a transformation that can be represented as the composition of a rigid Euclidean motion and a rescaling of RP.
Euclidean Motions and Isometries
Bya Euclidean motion of RP we shall mean a transformation h: RP Jo RP that can be written as the composition of a special orthogonal transformation and a translation of RP. The class Euc(p) of Euclidean motions of RP is a group and is commutative only for the case where p = 1. The group of Euclidean motions allows us to define the concept of congruence between subsets of RP. Two subsets Al and A z of RP are said to be congruent if there exists a Euclidean motion h E Euc(p) such that h(AI) = A z , or equivalently hI(A z ) = AI' The concept of congruence between sets forms the basis for Euclidean geometry, which involves the investigation of the geometric properties of subsets of Euclidean space RP. A property of a subset is said to be a geometric property if it is shared by any subset that is congruent to it. The definition of aEuclidean motion of RP can be generalized to arbitrary metric spaces. A metric space M is a seton which a metric d(x, y) is defined, where d satisfies the abstract properties (i), (ii), and (iii) of Problem 5 in Chapter 1.
2.1.9
2.1 Transformations on Euclidean Space
Background Concepts and Definitions
\
Once again, it can be checked that the class of similarity transformations forms a group under composition. Henceforth, we shall denote the class of similarity transformations of RP by Sim(p). The group of similarity transformations has a special representation when p = 1,2. In these cases, additional algebraic structure is available from multiplication of real and complex numbers respectively. In the latter case, we can again identify R Z with the complex plane C. Then we can write transformations in Sim(l) and Sim(2) in the form X Jo ax + b, where a f= 0 and b are arbitrary elements of R or C in the. respective dimensions. Multiplication and addition are the usual algebraic operations. Just as the group of Euclidean motions leads to the concept of congruence between sets, so the group of similarity transformations leads to the concept of similar sets. Definition 2.1.5. Two subsets Al and A z of RP are said to be similar or to have the same shape if there exists a similarity transformation hE Sim(p) such that h(AI) = A z or equivalently if hI(A z) = AI. If Al and A z are similar, then we shall write Al rv A z. We shall also be concerned with labeled figures or sets. For example, a triangle is often labeled at its vertices in Euclidean geometry in order to compare corresponding points on different triangles or simply to clarify a construction. The definitions of congruent and similar sets have obvious extensions to labeled sets, provided the labels correspond. Definition 2.1.6. We shall say that two correspondingly labeled sets have the same shape if one set can be transformed by a similarity transformation to the other set in such a way that labeled points are mapped to the corresponding points of the other figure. For example, two triangles XIXZX3 and YIYZY3 have the same shape if the angle at vertex Xj equals the angle at Yj for j = 1, 2, and 3. While the distinction between labeled and unlabeled sets can be regarded as a mathematical convenience in defining shapes, it is a more substantial distinction for the comparison of shape differences, as we noted in Chapter 1. An attempt to discover the shape differences between sets will typically involve a matching of the sets to determine how differences in the coordinates of corresponding points can be explained through similarity transformations. Any residual differences that cannot be explained through similarity transformations can be understood to be due to differences in shape. The problem of constructing .an appropriate correspondence between unlabeled sets (or unparametrized sets in general) is the problem of homology,
32
2. Background Concepts and Definitions
2.1 Transformations on Euclidean Space
A case that will be of particular interest to us occurs when p = q and A is of full rank. In this case, r is a square diagonal matrix whose diagonal elements are the singular values. The singular value decomposition allows us to represent a matrix in diagonal form, with IlJ and 1lJ' serving to provide a change of coordinate systems for the purpose. The singular value decomposition has an important geometric interpretation that will be of use in the next chapter. Suppose x is a 2 x 1 column vector and that A is a 2 x 2 matrix of full rank. Under the linear transformation x > Ax the unit circle in the plane. R 2 is mapped to an ellipse. The lengths of the semimajor and semiminor axes of this ellipse are seen from equation (2.3) to be the singular values of A. This geometric interpretation generalizes into higher dimensions. A p x p matrix A of full rank will have p singular values. If x is a p x 1 column vector, then x , Ax will map the unit sphere in RP to an ellipsoid with p principal axes. The singular values of A can be seen to be one half the lengths of the principal axes of the ellipsoid.
2.1.6
2.1.7 Groups of Transformations
';
(2.7) Suppose h is a 11 function that maps RP onto itself. We shall let h I denote the inverse function, where hl(y) = x whenever h(x) = y. There is nothing special about RP in these definitions, as RP can be replaced by CP or any other set.
Definition 2.1.1. A nonempty collection H = {h} of 11 transformations on a set is said to be a group provided that it is closed under composition and inversion of transformations.
The inner product between two elements, x (Yl,''''YP) of RP is defined as
In order for a nonempty collection H of transformations on a set to be a group, it is necessary and sufficient that for any h l , h 2 in H the transformation hI 0 h 2 I be in H. Setting h l = h2 we see that the identity transformation e is always an element of H.
P,
=
LXiYj
(2.4)
j=1
Its complex counterpart for CP is called the Hermitian inner product. We encountered the Hermitian inner product in Chapter 1 when defining the distance between two shapes in formula (1.21). We define the Hermitian inner product between two vectors x = (Xl> ..., x p) and Y = (Yl' ..., Yp) of complex coordinates to be
Definition 2.1.2. Two transformations h l and h2 are said to commute when h 20h l = h loh2 . We say that a group H is commutative or Abelian provided that any two elements of H commute.
p
«x,y» =
LXjyj
(2.5)
j=l
where Y~ is the complex conjugate of Yj E C. Under the identification of C with JR 2 the inner product on R 2 p can be defined from the Hermitian inner product on CP by noting that = lR « .,. ». Orthogonal transformations can be characterized as linear transformations that preserve inner products. Thus if A = (Aj k ) is an orthogonal matrix, then representing x, Y E RP as column vectors, wehave
< Ax, Ay > = < x, Y >
The classes O(p), SO(p), and their complex analogs U(p) and SU(p) are classes of transformations of a space to itself. now summarize some definitions and properties of groups, of which these classes are examples. Let hI and h2 be any two transformations from RP to RP. By the composition of b: and h 2 we shall mean the function h 2 0 hI from RP to RP defined by
We
Inner Products
< X,Y >
33
(2.6)
for all X and Y in RP. Similarly, Hermitian inner products on CP are preserved under unitary transformations.
.,.,
By the center of a group H we mean the set of elements of H that commute with every other element of H. Obviously, a group H is commutative if and only if the center of H is H itself. The class of orthogonal matrices is closed under matrix multiplication as well as matrix inversion. Similarly, the class O(p) of orthogonal transformations is closed under function composition and function inversion. Thus the class of orthogonal transformations is a group, that is commutative only for the cases where p = 1,2. The class SO(p) of special orthogonal transformations is a subgroup of the group of orthogonal transformations. That is, it is a subset of O(p) that is a group in its own right, being also closed under composition and inversion. When p = 1 this subgroup is the trivial group consisting of the identity transformation alone. Similar results hold true for the class of unitary transformations of CPo The class U(p) is also a group, containing the subgroup SU(p).
34
2.1.8
2.
Euclidean Motions and Isometries
By a Euclidean motion of RP we shall mean a transformation h: RP :> RP that can be written as the composition of a special orthogonal transformation and a translation of RP. The class Euc(p) ofEuclidean motions of RP is a group and is commutative only for the case where p = 1. The group of Euclidean motions allows us to define the concept of congruence between subsets of RP. Two subsets Al and A z of RP are said to be congruent if there exists a Euclidean motion h E Euc(p) such that heAl) = A z , or equivalently hI(A z ) = AI. The concept of congruence between sets forms the basis for Euclidean geometry, which involves the investigation of the geometric properties of subsets of Euclidean space RP. A property of a subset is said to be a geometric property if it is shared by any subset that is' congruent to it. The definition of a Euclidean motion of RP can be generalized to arbitrary metric spaces. A metric space M is a seton which a metric d(x, y) is defined, where d satisfies the abstract properties (i), (ii), and (iii) of Problem 5 in Chapter 1. Definition 2.1.3. A 11 correspondence h: M :> N between metric spaces is said to be an isometry if d(x, y) = d[h(x),h(y)] for all x, y E M. Two metric spaces are said to be isometric if there is an isometry mapping one to the other. When M and N ,are isometric, we shall write M ~ N.
In particular, the class of all isometries from M to itself shall be denoted Iso(M). It is immediate that the identity transformation on M is an isometry, and it can be checked that the transformations of Iso(M) form a group. On RP, for example, the class of Euclidean motions Euc(p) forms a subgroup of Iso(RP). This subgroup is a proper subgroup, because the Eucidean motions of RP do not inchidereflections through a (p  1)dimensional hyperplane. We may also speak of a linear isometry between vector spaces. Definition 2.1.4. A linear transformation of full rank between two vector spaces is said to be a linear isometry if it preserves the lengths of vectors.
Clearly, an orthogonal rotation of RP is an example of a linear isometry from RP to itself.
2.1.9
2.1 Transformations on Euclidean Space
Background Concepts and Definitions
Similarity Transformations and the Shape of Sets
Let (Xl, ...,xp ) be an element of RP. A transformation (Xl, ... ,X p ) :> (AXI, ..., AXp ) , where >. > 0, is said to be an isotropic rescaling or simply a rescaling of RP. By a shapepreserving transformation or a similarity trans
'"'H •
:> .
'F. 'j. '
"'.
35
formation of RP, we shall mean a transformation that can be represented as the composition of a rigid Euclidean motion and a rescaling of ~p. Once again, it can be checked that the class of similarity transformations forms a group under composition. Henceforth, we shall denote the class of similarity transformations of RP by Sim(p). The group of similarity transformations has a special representation when p = 1,2. In these cases, additional algebraic structure is available from multiplication of real and complex numbers respectively. In the latter case, we can again identify R Z with the complex plane C. Then we can write transformations in Sim(l) and Sim(2) in the form X :> ax + b, where a f= 0 and b are arbitrary elements of R or C in the respective dimensions. Multiplication and addition are the usual algebraic operations. Just as the group of Euclidean motions leads to the concept of congruence between sets, so the group of similarity transformations leads to the concept of similar sets. Definition 2.1.5. Two subsets Al and A z of RP are said to be similar or to have the same shape if there exists a similarity transformation hE Sim(p) such that heAl) = A z or equivalently if hI(Az) = AI' If Al and A z are similar, then we shall write Al rv A z.
We shall also be concerned with labeled figures or sets. For example, a .triangle is often labeled at its vertices in Euclidean geometry in order to compare corresponding points on different triangles or simply to clarify a construction. The definitions of congruent and similar sets have obvious extensions to labeled sets, provided the labels correspond. Definition 2.1.6. We shall say that two correspondingly labeled sets have the same shape if one set can be transformed by a similarity transformation to the other set in such a way that labeled points are mapped to the corresponding points of the other figure.
For example, two triangles XlXZX3 and YlYZY3 have the same shape if the angle at vertex Xj equals the angle at Yj for j = 1, 2, and 3. While the distinction between labeled and unlabeled sets can be regarded as a mathematical convenience in defining shapes, it is a more substantial distinction for the comparison of shape differences, as we noted in Chapter 1. An attempt to discover the shape differences between sets will typically involve a matching of the sets to determine how differences in the coordinates of corresponding points can be explained through similarity transformations. Any residual differences that cannot be explained through similarity transformations can be understood to be due to differences in shape. The problem of constructing an appropriate correspondence between unlabeled sets (or unparametrized sets in general) is the problem of homology,
36
2.2 Differential Geometry
2. Background Concepts and Definitions
discussed in Section 1.5. ]
2.2 2.2.1
Differential Geometry Homeomorphisms and DijJeomorphisms of Euclidean Space
Let h: U  t V be a continuous function between two open sets U c RP and V c Rq. Let us write (Yl,Y2, ..., Yq) = h(Xl' X2, ..., x p). We say that h is a smooth, or differentiable, mapping on U provided that h possesses finite partial derivatives 8Yk/8xj for all j = 1, ... ,p and all k = I, ... , q. If all these partial derivatives are continuous functions, then we say that h is a C1function on U. This definition can be extended to higher order derivatives. We say that h is a C function on U for any r = 1,2, ... if h has continuous partial derivatives , 8Tl+T2+ +TpYk (2.8) 8X~' ax~2
ax~p
for all k = 1, ... , q and all nonnegative integers rl, r2, ... , r p such that rl +r2+...+rp ~ r. Clearly, any function that is cr on U is a CSfunction for any s < r, If h is a CTfunction for all r 2: 1, then we say that h is a Coofunction. By convention, CDfunctions are understood to be the class of continuous functions on U. Associated with any smooth function h: U  t V and any point x = (Xl> .." x p ) in U is the Jacobian matrix. This is the matrix of partial derivatives
A
It defines a linear transformation u
(2.9)
t
Au where u is a p x 1 column
vector. This linear transformation
(2.10) is called the derivative of h at x. The Jacobian matrix can be regarded as a coordinate representation of the derivative of h. The derivative Dh. is the second term in the Taylor approximation to the function h at x, namely (2.11) h(x + u) = h(x) + ('Oh)x(u) + o(llulD
37
When p = q, the Jacobian matrix becomes a p x p square matrix. The determinant (.Jh)x = det(A) (2.12) is simply called the Jacobian of h at x, at the risk of some confusion. Note that the Jacobian matrix is a matrix valued function at each point x E U while the Jacobian at x is a real valued function. The Jacobian measures the rate of change of volume induced by the transformation x  t h(x) locally around x. Suppose that p = q and that h is a 11 correspondence from U to V. Then h is said to be a homeomorphism from U to V provided that both hand h 1 are continuous. When a homeomorphism can be established between U and V we say that U and V are homeomorphic. A homeomorphism h is called a CTdiffeomorphism between U and V if both hand h 1 are Cfunctions. We will normally refer to a coo_ diffeomorphism simply as a diffeomorphism. When a diffeomorphism can be established between U and V we shall say that U and V are diffeomorphic.
2.2.2
Topological Spaces
The properties of continuity and differentiability on RP can be abstracted to more general sets, leading to the concepts of the topological space, the topological manifold, and the differential manifold. Suppose M is a set endowed with a collection of subsets U = {U}. We say that U is a topology on the set M provided that (i) the empty set and M itself are both elements of U, (ii) any arbitrary union of elements of U is an element of U, and (iii) any finite intersection of elements of U is an element of U. The set M, endowed with a topology, is called a topological space, and the elements of U are called the open sets of M. A subset of M is said to be closed if its complement is open. The standard example of a topological space, which we have already considered, is when M is Euclidean space RP and U is the class of open sets of RP. Let M and N be topological spaces endowed with topologies Ul and U2 respectively. A function h: M  t N is said to be continuous if h1(U) E U1 for all U E U2 . If h is both 11 and onto, then we say that h is a homeomorphism provided that both hand h 1 are continuous. In RP a subset that is both closed and bounded has the property of compactness. This can be generalized to an arbitrary topological space. A subset A of a topological space M is said to be compact if every collection of open sets whose union contains A has a finite sub collection whose union also contains A. The HeineBorel theorem states that a subset of RP is compact if and only if it is closed and bounded. For our purposes in this and subsequent chapters, only a few properties of compactness will be used. Important among these properties is the fact that the continuous image of
38
2. Background Concepts and Definitions
a compact set is compact. More specifically,tLMand N are topological . spaces and h: M + N is a continuous function then h(A) iscompact for all compact subsets A C M.
2.2.3 Introduction to Manifolds A manifold is a generalization of our understanding ofacurved surface in three dimensions. We usually think of a curved surface as a subset of threedimensional Euclidean space R 3 that inherits its geometric properties from the geometric structure of the Euclidean space in which it lies. The representation of a space as a subset of another space is formally called an embedding. However, our intuition, being limited to objects in dimensionsless than or, equal to three, has trouble visualizing curvature of sets or spaces that cannot be embedded in threedimensional Euclidean space. The formal definition of a differential manifold has no such constraint. As much of calculus involves local constructions, differential manifolds, which locally resemble Euclidean space, become a natural domain for operations such as taking a gradient of a function, calculating tangent vectors, and other constructions from multivariable calculus. Examples of differential manifolds are common. A torus (the surface of a doughnut) is a differential manifold, as is a sphere or a flat plane. Some very small twodimensional being situatedin a torus would have trouble distinguishing the space around it from the space of a twodimensional sphere or a plane. This is because curved surfaces look approximately flat when viewed over a small region. The immediate vicinity of the being provides local information about the surface but little in the way of information about global properties of the surface that distinguish spheres from tori. To find global information, the being would have the walk around both surfaces and be very careful to 'check angles and distances. If the being were nearsighted and could not check distances and angles, then its examination of the local vicinity, or neighborhood, would fail to detect any local distortions due to the curvature of the surface, It might then conclude that the surrounding space was Euclidean, or flat, in nature. This is what we mean when we say that a differential manifold 10Qks locally like RP. Our twodimensional being might well consult an atlas to find its way around the geography of these twodimensional worlds. Weare used to seeing the surface of the Earth displayed in an atlas. However, we know that because the Earth is a sphere, we cannot get all points plotted on a single page or chart without tearing the picture and destroying the natural continuity between neighboring points. Just as a portion of the surface of the Earth can be described by a chart, so a portion of a differential manifold is described by a chart, here understood in a mathematicalsense. Just as a single page of an atlas cannot cover the entire surface of the Earth without disrupting continuity, so a single chart cannot usually cover the entire region of a differential manifold. The mathematical charts used to
2.2 Differential Geometry
39
describe a manifold must also be collected together into an atlas. Of course, such charts, if they cover the manifold, will overlap in places. Thus they are not arbitrarily related, but must, in a certain sense, describe the same smoothness on the region of overlap. If the same town appeared on twodifferent pages of a geographical atlas, we would expect the local descriptions on the two pages to be compatible, even if not identical. On a differential manifold, that notion of compatibility is described using a diffeomorphism.
2.2·4
Topological and Differential Manifolds
Let MP be a topological space with a collection of open subsets
{U", : aEA}
(2.13)
such that (2.14)
and a collection of functions (2.15)
that are all homeomorphisms onto the open subsets h(U",) of RP. Note that we do not assume {U'" : a E A} is the entire topology on MP. Then we say that the functions c'" are charts on MP provided that (2.16) is a homeomorphism from c'"(U'" n U{3) to c{3 (Uo n U{3) for all a and f3 in A. See Figure 2.1. We can think of the charts {c",}"'EA as providing local coordinate systems on MP. Formula (2.16) provides a patching criterion, telling us that these different coordinate systems can be glued together in a topologically consistent way. Definition 2.2.1. The collection of subsets {U"'}"'EA with the charts {c",}"'EA is said to form an atlas on MP. The set MP toqeiher with its atlas {(U""c",): a E A} is called a topological manifold of dimension p. A subset V C MP is open if c",(V n U",) is an open subset of RP for . every a E A. This definition formalizes our basic understanding that a topological manifold is a space that is locally homeomorphic to Euclidean space. Definition 2.2.2. If the functions C{3 0 C;:;l in (2.16) are also required to be Cr diffeomorphisms then the topological manifold MP is said to be a Cr differential manifold.
36
2.
2.2 Differential Geometry
Background Concepts and Definitions
37
discussed in Section 1.5.
When p = q, the Jacobian matrix becomes a p x p square matrix. The determinant (.Jh)x = det(A) (2.12)
2.2
is simply called the Jacobian of h at x, at the risk of some confusion. Note that the Jacobian matrix is a matrix valued function at each point x E U while the Jacobian at x is a real valued function. The Jacobian measures the rate of change of volume induced by the transformation X 4 hex) locally around x. Suppose that p = q and that h is a 11 correspondence from U to V. Then h is said to be a homeomorphism from U to V provided that both hand h l are continuous. When a homeomorphism can be established between U and V we say that U and V are homeomorphic. A homeomorphism h is called a C diffeomorphism between U and V if both hand h l are Cfunctions. We will normally refer to a coo _ diffeomorphism simply as a diffeomorphism. When a diffeomorphism can be established between U and V we shall say that U and V are diffeomorphic.
Differential Geometry
2.2.1
Homeomorphisms and Diffeomorphisms of Euclidean Space Let h: U > V be a continuous function between two open sets U c RP and V c Rq. Let us write (Yl,Y2, ...,Yq) = h(Xl,X2,''''Xp). We say that h is a smooth, or differentiable, mapping on U provided that h possesses finite partial derivatives 8Yk/8xj for all j = 1, ... ,p and all k = 1, ..., q. If all these partial derivatives are continuous functions, then we say that h is a Clfunction on U. This definition can be extended to higher order derivatives. We say that h is a Crfunction on U for any r = 1,2, ... if h has continuous partial derivatives , {jT! +r2+ +rpYk (2.8) 8x~! ax~2
'.~
.~
ax~p
2.2.2
for all k = 1, ..., q and all nonnegative integers rl, r2, ... , r p such that rl +r2+ ... +rp ~ r. Clearly, any function that is cr on U is a CSfunction for any s < r. If h is a crfu.nction for all r ~ 1, then we say that h is a COOfunction. By convention, CDfunctions are understood to be the class of continuous functions on U. Associated with any smooth function h: U 4 V and any point x = (Xl, ..., xp) in U is the Jacobian matrix. This is the matrix of partial derivatives !!J& !!J&
The properties of continuity and differentiability on RP can be abstracted to more general sets, leading to the concepts of the topological space, the topological manifold, and the differential manifold. Suppose M is a set endowed with a collection of subsets U = {U}. We say that U is a topology on the set M provided that (i) the empty set and M itself are both elements of U, (ii) any arbitrary union of elements of U is an element of U, and (iii) any finite intersection of elements of U is an element of U. The set M, endowed with a topology, is called a topological space, and the elements of U are called the open sets of M. A subset of M is said to be closed if its complement is open. The standard example of a topological space, which we have already considered, is when M is Euclidean space RP and U is the class of open sets of RP. Let M and N be topological spaces endowed with topologies Ul and U2 respectively. A function h: M 4 N is said to be continuous if hl(U) E Ul for all U E U2' If h is both 11 and onto, then we say that h is a homeomorphism provided that both hand h 1 are continuous. In RP a subset that is both closed and bounded has the property of compactness. This can be generalized to an arbitrary topological space. A subset A of a topological space M is said to be compact if every collection of open sets whose union contains A has a finite subcollection whose union also contains A. The HeineBorel theorem states that a subset of RP is compact if and only if it is closed and bounded. For our purposes in this and subsequent chapters, only a few properties of compactness will be used. Important among these properties is the fact that the continuous image of
axp
aX!
A
(2.9)
!!J!...
o
/
o c a (U an u~
FIGURE 2.1. Charts on a manifo/4. A chart provides a coordinate system on a manifold. In order to ensure that the coordinate systems are consistent wit/!, each other, a patching criterion is required on the sets of the manifold where the coordinate systems overlap. For the figure shown, the patching criterion requires that cl3 0 C;;l be a diffeomorphism between subsets of RP. A set of compatible charts that cover the manifold is called an atlas. In RP it is often useful to change coordinate systems for the convenience of calculations. The same is true for differential manifolds. As it is the charts that provide coordinates for points in the manifold, a change of coordinate systems about a point x E MP is simply a change in the choice of the chart c'" that provides coordinates for x. If the change in coordinates is to be compatible with the differential, structure defined on MP, then the new chart C{3 will need to satisfy the patching crit~rion above. Such a criterion will automatically be satisfied if the chart CI3 belongs to the atlas on MP. II owever, the new chart is not required to belong to the atlas. If the new chart satisfies the patching criterion, it can be included in the atlas, and thereby enlarge the atlas.
is differentiable. Similarly, we will say that h is a cr function provided that the function defined in formula (2.18) is a ·Crfunction. If p = q and h is 11 and onto, then h is called a Crdiffeomorphism provided that h and h l are cr. Once again, when h is a COOdiffeomorphism, then we shall simply refer to h as a diffeomorphism. When a crdiffeomorphism can be established between two manifolds MP and NP then MP and NP are said to be Cr diffeomorphic. If r = 00 then we shall simply say that MP and NP are diffeomorphic. Atlases provide coordinate systems for manifolds. For example, if x is a point in Ua then the coordinates of ca(x) in the Euclidean space RP can be used to locate the point. Unfortunately, there is usually no single chart that can provide a nondegenerate coordinate system simultaneously for the entire manifold, as charts have to be patched together to cover the manifold. However, in many cases, the points of degeneracy of coordinate systems introduced by charts need not be a hindrance to calculations. For this reason, we often suppress the chart notation, and say that point x has coordinates (Xl, X2, ... , xp) rather than the more precise statement that these coordinates belong to ca(x). The intrinsic properties of a manifold are those that are invariant under a change of coordinates that is compatible with the differential structure, as explained in Figure 2.1. On the other hand, those properties that are dependent upon the coordinate system are called extrinsic properties of the manifold. As we defined a differential manifold to be a space that is locally diffeomorphic to Euclidean space, it is not surprising that Euclidean space RP turns out to be a differential manifold. To do this, we make the atlas consist of a single chart, with U = RP and C = e, where e is the identity transformation from RP to RP. With this construction, it becomes a routine matter to check that RP satisfies the definition of a differential manifold.
42
2. Background Concepts and Definitions
2.2 Differential Geometry
43
\\
2.2.5 An Introduction to Tangent Vectors Let us return to our intuitive example of a differential manifold, namely a surface embedded in Euclidean space R 3 . A typical way in which a surface can be defined is as the solution set to an equation of the form (2.19)
where h is a real valued function defined on R 3 . Let us denote this surface by M 2 as shown in Figure 2.2. Suppose that x = (Xl, X2, X3) is a point on this surface. Now if the gradient vector oh oh oh ) Vh(x) = ( ~,~, ~ UXI UX2 UX3
(2.20)
is nonvanishing, it will point in a direction perpendicular to the surface. Tangent vectors to the surface at the point X will then be those vectors in R 3 that are orthogonal to this normal vector. The set of all vectors that are tangent to the surface at X is said to be the tangent space of the surface at x. Thus a vector v = (VI, V2, V3) is a tangent vector to the surface at the point x if and only if
a
(2.21)
When we turn to general differential manifolds this construction unfortunately does not generalize. Nevertheless, the space of tangent vectors can be defined in a more abstract sense, despite the fact that anormal vector to a surface is a property of the embedding .in R 3 and not intrinsic to . the differential geometry of that surface. A key insight in generalizing the concept of a tangent vector is to note that ona surface, the tangent vectors at a point x can be placed in 11 correspondence with equivalence classes of paths through x, which we shall now consider. Let Xo be a point on the surface M 2 • Now let (2.22)
be a path in the surface passing through a point Xo at t = 0 and defined for values of t in some open interval (E, E). For each t, define the vector i:(t) by
i:(t)
=
dx(t) dt
= .(.. dXI(t) dt
'
x(t)
dX2(t) dX3(t)) dt ' dt
FIGURE 2.2. Tangent and normal vectors to a surface. At any point on a surface, the tangent vectors to the surface are perpendicular to a normal vector that is the gmdient of the defining equation.
point. However, all paths through Xo having the same tangent vector at Xo form an equivalence class of paths. It is this equivalence class that we will formally identify with the tangent vector at Xo in the next section. We close this section by considering how tangent vectors to a surface can be used to represent infinitesimal displacements of points within the surface. Consider Figure 2.3. Along a smooth path in a surface, position two points x and y. From the point x draw a vector in R 3 out to y. This vector is called a secant vector because it points along a secant line segment whose endpoints are the two points x and y in the surface. Secant vectors point in the direction of the displacement from x to y, but are represented in the Euclidean space R3 rather than the surface itself. When the displacement from x to y becomes infinitesimally small, then the secant vector in limiting form becomes a tangent vector to the surface. Thus we can write dx = i:(t) dt where dx = x(t + dt)  x(t) and i:(t) is, once again, the tangent to the curve at x = x(t). Thus the length ds of the displacement dx is
(2.23)
Then it can be seen that the vectori:(O) is a tangent vector to the surface at the point xo. See Figure 2.2. Thus every smooth path through xodefines a tangent vector at xo. This tangent vector is not unique to the path,as there exist many paths through Xo having the same tangent vector at that
ds = 11i:(t)11 dt
2.2.6
(2.24)
Tangent Vectors and Tangent Spaces
Henceforth, we shall assume that MP is a differential manifold. Let x(t) and y(t) be two smooth paths in MP passing through a common point Xo
44
2.2 Differential Geometry
2. Background Concepts and Definitions
45
;«1)
FIGURE 2.3. Secant vectors to a surface. In the limit, as the displacement between points becomes infinitesimal, the secant vectors converge to a tangent vector.
at t = 0, say. Let us suppose that a coordinate system has been constructed by a chart (Uc':i; to be the equivalence class of paths tangent at t = 0 to the path with coordinates (2.30) Scalar multiplication can also be shown to be well defined. Note that we
46
2.
Background Concepts and Definitions
2.2
can add tangent vectors at the same point Xo but cannot add tangent vectors that are tangent to the manifold at different points.
2.2:7
Differential Geometry
47
Metric Tensors and Riemannian Manifolds
Suppo~~ that
Definition 2.2.4. The vector space of all tangent vectors to the manifold MP at a given point x E MP is called the tangent space at x and is denoted by Tx(MP). g(X)
The tangent space Tx(MP) can be shown to have the same dimension as the manifold MP. So Tx(MP) is linearly isomorphic to Euclidean space W. . Within Tx(MP) it is possible to construct a set of basis vectors as follows: For each j = 1, ... ,p consider the path
is a positive definite symmetric matrix for all x E MP. Then g(x) defines an inner product on Tx(MP) as follows. Consider two tangent vectors in Tx(MP), namely E j aj (x)Oj (x) and Ek bk(x)Ok(X). Then we define the inner product of these tangent vectors to be
(2.31) defined in a neighborhood of x = (Xl, ..., x p ) around t = O. These paths pass through the point x at t = 0 and follow the axes of the coordinate system about x. For each j = 1, ... ,n we define OJ(x) E Tx(MP) to be the tangent vector to the path defined by formula (2.31) at the point x where t = O. The tangent vectors Ol(X), 8 2(x ), ..., op(x) collectively form a basis for the tangent space Tx(MP). That is, any tangent vector in TAMP) can be written as
P 'r;,
(2.32)
where each aj is a real valued function of x E MP. For example, we can write LXj(t) OJ[x(t)]
>
P
= LLgjk(x)aj(x)bk(X)
k=l
(2.36)
j=l k=l
< L ajoj , L bkOk >
P
=
P
P
Laj(x)oj(x), Lbk(X)Ok(X)
This notation is cumbersome if used on a regular basis. We shall suppose that gjk is a smoothly varying function in x across the manifold and shall suppress the x, both in gjk and the tangent vectors. Thus we can also write this in more compact form as
j=l
x(t)
=
(2.40)
be the length of the vector x(t). The inner product generated by the metric tensor can be calculated using formula (2.37). So we can write P
P
L L 9jk(t)Xj(t)Xk(t)
(2.41)
j=lk=l
where 9jk(t) is the value of the metric tensor at x(t). Suppose t undergoes a small increment to t + dt. Then, as in formula (2.24), the length ds of the path segment from x(t) to x(t + dt) is (2.42)
ds = I(t) dt
Therefore we can write the length of the path x(.) from t = to to t = tl as L
=
i
t,
to
ds
=
itl
"I(t) dt .
(2.43)
to
It should be noted that not only does the metric tensor determine the lengths of arcs, but the metric tensor is also itself determined by the arc length. That is, if ds can be calculated for any increment of a path from x(t) to x(t + dt) then there is at most one metric tensor 9 that is compatible with this definition. In some cases, we shall determine the structure of a Riemannian manifold by calculating the arc length function ds. Roughly speaking, a geodesic path on a Riemannian manifold is the path between two points that has shortest length. This definition is a bit too narrow to work but serves for the basic intuition. More correctly, we can say that a geodesic x(t) is a path in a Riemannian manifold that is locally shortest. This means that the path can be broken up into pieces such that
FIGURE 2.5. The geodesic path on a manifold displayed as the path of locally shortest length. On the sphere we see a variety of paths between two points. The shortest path is a geodesic between the two points, which in this case is an arc of a great circle of the sphere. the paths connecting the endpoints of the pieces are all the shortest paths. This definition does not require that the endpoints of the path x(t) be specified in order to determine whether it is a geodesic: the property can be investigated locally along the path. See Figure 2.5. In Euclidean space RP, the shortest distance between two points is, of course, a straight line. Thus the geodesic paths of a manifold can be regarded as the analogs of straight lines for spaces that are not flat. We can find formulas for geodesic paths by applying the calculus of variations to the arc length formula (2.43) above. To find a condition to ensure that this path is minimal in length, we consider a perturbation of the path along a coordinate. If the integral is minimized then its derivative with respect to this perturbation is zero. This leads to the following set of equations from the calculus of variations. The path is a geodesic provided the EulerLagrange equations are satisfied, namely that (2.44) for all j = 1, 2, ... , p. To interpret the partial derivatives in this formula, note that for fixed t the expression I(t) depends upon Xl (t), ..., xp(t) and Xl(t), ... ,xp(t) . Variation in the position coordinates Xj(t) is suppressed in the notation, but arises from the metric tensor 9 in formula (2.41), which is a function of (Xl, ... , x p ) . The partial derivatives are understood to be the partial derivatives in each of these 2p variables holding the other 2p  1 variables fixed. The example in Section 2.2.10 below shows how to interpret this formula in RP with the usual coordinate system. Problems 5 and 6 at the end of the chapter ask the reader to check the equations for various settings.
=~
50
2. Background Concepts and Definitions
2.2 Differential Geometry
51
Having defined the concept of a geodesic path in a Riemannian manifold, we are in a position to define the concept of the geodesic distance between two points in the manifold.
ofthe curvature of the manifold.
Definition 2.2.6. Suppose that a Riemannian manifold MP is pathwise connected, in the sense that for any two points x, y E MP there exists a smooth path x(t) such that x(to) = x and x(t 1 ) = y. We define the geodesic distance from x to y to be the length of the shortest path from x to y.
We consider the geodesics on RP and check that they are straight lines. The usual Cartesian coordinates are used so that the atlas consists of a single chart (RP, e), where e: RP + RP is the identity map. The metric tensor g is the p x p identity matrix. Consider a smooth path x(t) in RP. Then
2.2.10
Example
P
With this definition, a pathwise connected Riemannian manifold becomes a metric space, as was defined in Problem 5 of Chapter 1. It can be shown that the path of shortest length is a geodesic in MP. However, the converse does not hold. The length of a geodesic path from x to y can be strictly greater than the geodesic distance from x to y. This can easily be seen by considering the fact that on a sphere any great circle passing through two distinct points can be subdivided into two paths from one point to the other. Both of these paths are geodesics, but their lengths need not be equal. It is the smaller of these two lengths that is the geodesic distance from one point to the other.
2.2.9
Affine Connections
Closely related to the concept of a geodesic path is the concept of an affine connection. We noted earlier that the metric tensor allows us to compare the lengths and orientations of tangent vectors within a tangent space T x (MP). However, the metric tensor does not give us a direct method of comparing vectors in different tangent spaces, say Tx(MP) and Ty(MP). The way we would naturally think of doing this is to rigidly transport a vector from one place in the manifold to another. For example, we could draw a geodesic from x to y and move a vector along the geodesic so that its length remains constant and its angle with respect to the tangent vector of the geodesic path is also constant. A method for transporting tangent vectors is called an affine connection. The particular method just described using geodesics and the metric tensor is called the LeviCivita connection. A curious property of connections such as the LeviCivita connection is that when vectors are transported around the manifold along a sequence of geodesic paths, they can arrive back at their starting place with a different orientation from the one they started with. This is paradoxical when we recall that the method of transport associated with the LeviCivita connection requires that the orientation remain fixed with respect to the paths. However, the reader can try it on a sphere and observe this, moving a vector from the north pole to the equator, part way around the equator, and back to the north pole again. This change in orientation is a consequence
'Y =
LXJ(t)
(2.45)
j=l
The partial derivative 8'Y/8xj on the lefthand side of (2.44) can be computed directly from this formula by holding all other Xk constant for k I j. We obtain 8'Y Xj Xj (2.46) 8xj 'Y Ilxll This is a directional cosine of X. Thus the lefthand side measures how this directional cosine of the tangent vector along the path changes. On the righthand side of the EulerLagrange equations the partial derivatives are all zero because the metric tensor gjk in the formula for 'Y(t) is a constant function of position x(t). Therefore, we see that the EulerLagrange equations reduce to stating that the directional cosines of the path are constant. The path must therefore be a straight line.
2.2.11
Building New Manifolds From Old: Product Manifolds
Just as it is possible to build Euclidean spaces of arbitrarily high dimension by taking Cartesian products of R, so it is possible to build differential manifolds by taking Cartesian products of differential manifolds. Suppose MP and N? are differential manifolds of dimension p and q respectively. We can make MP x Nq into a differential manifold by using charts of the form CUe< x V,a , Ce< x c,a) (2.47) where (Uc",ce
N?
(2.54)
be 'a differentiable function, and suppose that x(t) is a smooth path in MP. Then h[x(t)] can be seen to be a smooth path in the manifold Nq. Differentiable mappings preserve tangency. For example, if Xo is any point on the path x(t), and if y(t) is a path in MP that is tangent to x(t) at xo, then h[y(t)] is tangent to h[x(t)] at the point h(xo) E N". It follows from this fact that h maps the equivalence class of paths in MP tangent to x(t) at Xo to the equivalence class of paths in Nq tangent to h[x(t)] at h(xo). But these equivalence classes are tangent vectors at Xo and h(xo) respectively. So this defines a mapping (2.55)
Definition 2.2.7. The mapping (Vh)x in formula (2.55) above is called the derivative of h at X E MP I and can be shown to be a linear transformation between the tangent spaces. We can also define (Vh)x directly using coordinates on the manifold. In terms of the coordinates (2.56)
(2.53) make a local smooth coordinate system of the right dimension on the submanifold. Using the coordinate system X = (Xl' X2, ... , X p ) we can set up the basis 8 1,cJz ,..., 8p for the tangent space Tx(MP). Among these basis vectors, the first q tangent vectors Eh, 82, ... , 8q form a basis for the tangent space Tx(Nq). Thus any tangent vector in Tx(Nq) can be written as 'L.:]=l aj(x)8j(x). If MP is a Riem~nnian manifold, then Nq can be made into a Riemannian manifold by inheriting the concept of arc length from MP. A geodesic path in Nq is simply a path of shortest length in MP among those constrained to lie wholly within Nq. If 9 is the metric tensor associated with the coordinate system (Xl, ... , x p ) then the induced metric tensor on N? is constructed as the q x q matrix consisting of the first q rows and columns of g.
2.2.13
Derivatives of Functions between Manifolds
In 2.2.1, we defined the derivative of a differentiable function h: U > V, where U and V are open sets of RP and Rq respectively. We shall now
suppose that we can write hex) as (2.57) Let ~l' ,8r: be the coordinate basis of Tx(MP), and correspondingly, , 8q be the coordinate basis for Th(x)(Nq). Then (Vh)x can let 8 1 , be expressed in terms of these basis vectors as . . (2.58) where bk
=
taj j=l
8h k 8xj
(2.59)
The expression can be seen to be left multiplication
a
>
Aa
(2.60)
where a = (al, ...,ap ) is the row vector of coefficients and A is the Jacobian matrix of the coordinate transformation from RP to Rq.
54
2.2 Differential Geometry
2. Background Concepts and Definitions
2.2.14
Again, we let 8 P denote the sphere of radius r
Example: The Sphere
55
= 1. An atlas
We finish this chapter with some examples of differential manifolds that will be useful in the next chapter. Examples of manifolds that are surfaces in R 3 (and one surface that is not) can be found in Problems 26 at the end of the chapter. . In R3, let 8 2(r ) , r » 0 be the set of all points z = (Xl, X2, X3) such
(2.61')
that (2.61) For notational simplicity, we typically let 8 2 denote the special case where 8 2 (r ) has canonical radius r = 1. The set 8 2 (r ) is called the 2sphere of radius r. We can put an atlas on 8 2 (r ) using the open sets Ul+, Ul  , and correspondingly the open sets U2+,U2 and U3+, U3, where Uj+ and Uj _ are the set of points of 8 2 (r ) with positive and negative xr coordinate respectively. To define a chart on Ul+ we set (2.62) Similarly, we define Cl(X) = (X2,X3) on Ul. Charts C2+,C2,C3+, and C3 on the other open sets are defined correspondingly. Although these coordinate systems establish 8 2 (r ) as a differential manifold, there are more charts than necessary. A minimum of two charts is necessary to define an appropriate atlas on 8 2 (r ) that corresponds to our intuitive understanding of the geometry of the sphere. For practical calculations, it is usually sufficient to set up a coordinate system through a single chart. These coordinates are the longitude (h and the colatitude (J2, defined so that the point
can be imposed on SP(r) in a manner similar to the 2sphere above. The jsphere s' is simply the unit circle. The usual geodesic distance betweentwo points of SP is the shorter of the two arcs of the great circle joining the points. This is simply the angle made between the two vectors from the origin to the two points. Thus if X and yare elements of 8 P C RP+l the geodesic distance from X to y is given by d(x,y) = cos"! « x,y » (2.68) where again is the usual inner product on RP+l. More generally, on the sphere SP (r), the geodesic distance from X to y is
d(x, y) = r cos"!
(r 2 < X, Y »
(2.69)
The Cartesian product 8 P x sq of two spheres SP and sq is a generalization of a torus, which becomes the special case when p = q = 1. Although the representation of the torus 8 1 x 8 1 is as a subset of R 4 this torus is well known to be diffeomorphic to a surface in R 3 that i~ the boundary of a doughnut. See Problem 2. However, the next example we shall consider is a twodimensional manifold or surface that cannot be represented as a subset of R 3 .
(2.63)
2.2.15
Example: Real Projective Spaces
has coordinates ((Jr, (J2). To impose the usual metric of great circle distance on 8 2 (r) we introduce the metric tensor 9 = (gjk) for the coordinate system ((Jr, (J2) where
In consider the set of all lines passing through the origin. Any such line can be represented as the set of scalar multiples
(2.64)
(2.70)
R3,
and (2.65) The offdiagonal elements g12 = g21 are set to zero. The geodesics of the manifold can be shown to be arcs of great circles. Extending to arbitrary dimensions is straightforward. In general, the psphere of radius r will be denoted 8 P (r ) and can be identified with the set of all points (Xl, X2, ... , x p ) in RP such that
X~
+
x~
+ ... +
x;
= r2
(2.66)
Definition 2.2.8. We call the set of such lines through the origin real projective 2space and symbolize it as RP2. As any line through the origin meets the unit sphere about the origin in exactly two antipodal points, it can be seen that real projective 2space is naturally identifiable with the set of all pairs of antipodal points on the unit sphere. See Figure 2.6. This representation is particularly useful in making Rp2
; I
52
/
2. Background Concepts and Definitions
those' of MP and N? so that (2.50) With this understanding, we can make MP x N? into a Riemannian manifold by putting the metric tensor elements as blocks down the main diagonal. If gAl is a metric tensor on MP and 9 N is a metric tensor on N? then an appropriate metric on MP x N" is (2.51)
2.2.12
Building New Manifolds Prom Old: Submanifolds
It is also possible to construct new manifolds by looking inside a manifold. Suppose N? is a subset of a differential manifold MI'. We say that N" is a qdimensional submanifold of MP for q < p if for every point y E Nq, there exists a smooth coordinate system x = (Xl, ..., Xl') on some open set U c MP containing y such that Un N?
= {x
E
U : xq+l
= Xq+2 = ... = xl' = O}
(2.52)
More informally we can say that a qdimensional submanifold of MP is a subset that is locally diffeomorphic to a linear subspace. The submanifold N? inherits a coordinate system from this construction. The coordinates
(2.53) make a local smooth coordinate system of the right dimension on the submanifold. Using the coordinate system x = (Xl,X2, ...,x p) we can set up the basis 01,{)2,.."Op for the tangent space Tx(Mp). Among these basis vectors, the first q tangent vectors 01, ~, ..., Oq form a basis for the tangent space Tx(Nq). Thus any tangent vector in TANq) 'can be written as 2:3=1 aj(x)oj(x). If MP is a Riemannian manifold, then N? can be made into a Riemannian manifold by inheriting the concept of arc length from MI'. A geodesic path in N" is simply a path of shortest length in MP among those constrained to lie wholly within N". If 9 is the metric tensor associated with the coordinate system (Xl, ..., Xl') then the induced metric tensor on N? is constructed as the q x q matrix consisting of the first q rows and columns of g.
2.2.13
2.2 Differential Geometry
/
Derivatives of Functions between Manifolds
In 2.2.1, we defined the derivative of a differentiable function h: U > V, where U and V are open sets of RP and R? respectively. We shall now
53
extend our definition to the case where h is defined between differential manifolds. Let (2.54) I be 'a differentiable function, and suppose that x(t) is a smooth path in MI'. Then h[x(t)] can be seen to be a smooth path in the manifold N". Differentiable mappings preserve tangency. For example, if Xo is any point on the path x(t), and if yet) is a path in MP that is tangent to x(t) at xo, then h[y(t)] is tangent to h[x(t)] at the point h(xo) E N", It follows from this fact that h maps the equivalence class of paths in MP tangent to x(t) at Xo to the equivalence class of paths in N? tangent to h[x(t)] at h(xo). But these equivalence classes are tangent vectors at Xo and h(xo) respectively. So this defines a mapping
(2.55) Definition 2.2.7. The mapping (Dh)x in formula (2.55) above is called the derivative of h at X E MP, and can be shown to be a linear transformation between the tangent spaces. We can also define (Dh)x directly using coordinates on the manifold. In terms of the coordinates
X
= (Xl,
X2, ..., Xl')
(2.56)
suppose that we can write hex) as
(2.57) Let ~1' , Of be the coordinate basis of Tx(Mp), and correspondingly, let 0 1 , , Oq be the coordinate basis for Th(x)(Nq). Then (Dh)x can be expressed in terms of these basis vectors as
(2.58) where
I'
bk = Laj oh k j=l OXj
(2.59)
The expression can be seen to be left multiplication a
>
Aa
(2.60)
where a = (a 1, ... , ap ) is the row vector of coefficients and A is the Jacobian matrix of the coordinate transformation from RP to Rq.
54
2.2
.2. Background Concepts and Definitions
2.2.14
Example: The Sphere
DifferentialCeometry
55
Again, we let SP denote the sphere of radius r = 1. An atlas
We finish this chapter with some examples of differential manifolds that will be useful in the next chapter. Examples of manifolds that are surfaces in R 3 (and one surface that is not) can be found in Problems 26 at the end of the chapter. In R3, let S2(r) , r > 0 be the set of all points x = (XL,X2, X3) such that (2.61)
(2.67)
can be imposed on 8 P (r ) in a manner similar to the 2sphere above. The Isphere Sl is simply the unit circle. The usual geodesic distance betweehtwo points of SP is the shorter of the two arcs of the great circle joining the points. This is simply the angle made between the two vectors from. the origin to the two points. Thus if x and y are elements of SP c RP+1the geodesic distance from X to y is given by d(x, y) = cos l « X, y » (2.68)
For notational simplicity, we typically let S2denote the special case where S2(r) has canonical radius r = 1. The set S2(r) is called the 2sphere of radius r, We can put an atlas on S2(r) using the open sets Ul+, Ul  , and correspondingly the open sets U2+, U 2 and U3+, U 3, where Uj+ and Uj _ are the set of points of S2(r) with positive and negative Xjcoordinate respectively. To define a chart on Ul+ we set
where again is the usual inner product on RP+I. More generally, on the sphere SP(r), the geodesic distance from X to yis
(2.62) Similarly, we define Cl_(X) = (X2,X3) on Ul  . Charts £:2+,C2,C3+, and C3 on the other open sets are defined correspondingly. Although these coordinate systems establish S2(r) as a differential manifold, there are more charts than necessary. A minimum of two charts is necessary to define an appropriate atlas on S2(r) that corresponds to our intuitive understanding of the geometry of the sphere. For practical calculations, it is usually sufficient to set up a coordinate system through a single chart. These coordinates are the longitude fh and the colatitude fh, defined so that the point
d(x, y)
..~ l;
= r cos"! (r 2 < X, Y »
(2.69)
The Cartesian product SP x sq of two spheres SP and sq is a generalization ofa torus, which becomes the special case when p = q = 1. Although the representation of the torus 8 1 x Sl is as a subset of R 4 , this torus is well known to be diffeomorphic to a surface in R3 that is the boundary of a doughnut. See Problem 2. However, the next example we shall consider is a twodimensional manifold or surface that cannot be represented as a subset of R 3.
(2.63)
2.2.15
Example: Real Projective Spaces
has coordinates (01,0 2 ) , To impose the usual metric of great circle distance on S2 (r) we introduce the metric tensor 9 = (9jk) for the coordinate system (01 , O2 ) where
In R 3, consider the set of all lines passing through the origin. Any such line can be represented as the set of scalar multiples
(2.64)
(2.70)
and 922. = r
2
(2.65)
The offdiagonal elements g12 = g2l are set to zero. The geodesics of the manifold can be shown to be arcs of great circles. Extending to arbitrary dimensions is straightforward. In general, thepsphere of radius r will be denoted. SP(r) and can be identified with the set of all points (Xl, X2, ... , x p ) in RP such that
(2.66)
Definition 2.2.8. We call the set of such lines through the origin real projective 2space and symbolize it as RP2. As any line through the origin meets the unit sphere about the origin in exactly two antipodal points, it can be seen that real projective 2space is naturally identifiable with the set of all pairs of antipodal points on the unit sphere. See Figure 2.6. This representation is particularly useful in making Rp2
56
2. Background Concepts and Definitions \
\
2.2 DifferentialGeometry
57
011 Rp2 by defining the three open sets
, \
\
,
\
(2.76)
/ /
\
/
/
where Uj+ and Uj _ 2 Cl'i U1 > R to be
are as defined for the sphere 8 2 above. Define
(2.77) Similarly, we define
C2:
U2
>
R 2 to be
(2.78)
/ / /
,
(2.79)
\
\
\ \
This particular differential structure on Rp2 has the property that a function f: Rp2 > R is differentiable if and only if the function
\ \
FIGURE 2.6. Real projective 2space represented as the space of lines passing through the origin in 3dimensional Euclidean space. Each line defines a pair of antipodal points on the unit sphere. Therefore a 11 correspondence exists between such lines and pairs of antipodal points.
into a differential manifold. Note that there is a natural mapping (2.71) that maps any point of the unit sphere to the set of two antipodal points of which it is an element. This mapping is an example of a special type of differentiable function between manifolds called a covering mapping. The image under A of any point x = (Xl, X2, X3) E 8 2 is the pair of antipodal points (2.72) A(x) = {x, x} To make Rp2 into a differential manifold, we can modify the charts of Example 2.2.1. Note that
foA: 8 2
>
R
(2.80)
is differentiable. Another property of the covering mapping A is that its derivative (2.81) is a linear transformation of full rank, i.e., is onto, at all points x E 8 2 . This fact can be used to motivate a particular choice of metric tensor on RP2. As (VA)x maps onto TA(x) (RP2), we can write any element of this tangent space as (2.82) where 8 1 and 8 2 form a coordinate basis for T x(8 2 ) . Let 8; = (VA)x(8j ) for j = 1,2. Then 8~ and form a basis for T A(x)(R p 2). If in addition, we set (2.83)
a;
then (2.73) Similarly, we have (2.74) These three sets are open in the natural topology on RP2. For any real number a =I 0, let sgn(a) denote the sign of the number a. We construct charts (2.75)
(2.84) becomes a linear isometry between tangent spaces. With this metric tensor on Rp2, the covering map A maps the geodesic great circles of 8 2 to geodesic paths in Rp2 and A becomes a local isometry between Riemannian manifolds. That is, if x and yare points of 8 2 separated by a geodesic distance of less than 'ff /2 then the geodesic distance from x to y in 8 2 equals the geodesic distance from A(x) to A(y) in R P 2 . However, the two manifolds are not isometric because A is not 11.
58
2.2 DifferentialGeometry
2. Background Concepts and Definitions
59
Extensions, some straightforward and others more substantial, are possible.
tained by replacing the real coordinates of Euclidean space RP+l with complex coordinates. We shall encounter this space in the context of shape manifolds in Section 3.2 in the next chapter.
Definition 2.2.9. We define real projective pspace, denoted by Rpp, to be the space of lines through the origin in RP+l. This space can be interpreted as the set of antipodal pairs of points on the pdimensional unit sphere SP C RP+l .
2.2.16
The manifolds that we shall consider next will be written with complex coordinates in what follows. However, they can be understood as examples of the differential manifolds that we have been discussing up to now. This can be seen through the identification of R 2 with the complex plane C. The differential manifold CpP that we shall consider will have p complex dimensions or equivalently 2p real dimensions. It can be regarded as a collection of complex lines through the origin in Cp+l or as a collection of planes through the origin in R 2P+2. Note that the latter interpretation has to be made with some care after identifying R 2 with C. Every complex line through the origin of CP can be considered as a plane through the origin in R 2p+2. However, the converse is not true. Similarly, we saw earlier that every unitary transformation of CP is an orthogonal transformation of R 2p, without the converse holding. Let CpP be the collection of all complex lines
The constructions above generalize in a natural way. Again, the covering map A: SP > RpP establishes a local isometry between SP and RpP that is not an isometry. To visualize what this means, consider the case where p = 1. In this case, it can be shown that Rp l is a circle of radius 1/2. We can establish a local isometry between a circle (2.85) by noting that the covering map A wraps the circle 8 1 , whose circumference has length 21f twice around the circle Sl(1/2), whose circumference has length 1f. This is akin to winding a thread tightly twice around a spool and then joining the ends of the thread to form a loop. The change of radius by a factor of one half is a natural consequence of the fact that the covering mapping A is 2 to 1. A consequence of this construction is that the geodesic distance between two points A( x ) and A(y) in RpP is found to be the smaller of the two geodesic distances from d(x, y) and d(x, y) in SP. This can be used as a definition of the metric on RPP without reference to the metric tensor. Furthermore, the image under the mapping A of a geodesic great circle path in 8 P is a geodesic path in RpP. The projective spaces RpP have a role in the representation of shapes. In Section 1.4.1, we noted that the preshape of n landmarks along a line can be represented as a point on a sphere of dimensionn 2. Antipodal points on this sphere represent the preshapes of reflected configurations. For example, if Xl, X2, and X3 are three landmarks on the real line, then the preshapes of (Xl, X2, X3) and (Xl, X2, X3) are antipodal points on the circle. This can be seen in Figure 1.3, where the three pairs of antipodal points displayed are the preshapes of equally spaced landmarks. So the projective spaces Rpn2 are appropriate manifolds for representing the preshapes of aligned landmarks when the distinction between a configuration and its reflection is. ignored. This is particularly appropriate for the example in Section 1.4.1, wherea reflection along the line of alignment corresponds to a rotation by 180 0 in the plane in which the images lie. We can summarize this by saying that the shapes of n 2: 3 aligned landmarks in the plane can be naturally represented as elements of the real projective space Rpn2. A variant of real projective space, which we shall consider next, is ob
Example: Complex Projective Spaces
(2.86) found by taking a point (Zl, Z2, ... , Zp+l) E Cp+l distinct from the origin and drawing the complex line through this point and the origin (0,0, ...,0). Any such complex line intersects the sphere (2.87) in a great circle. These circles partition the sphere, so that any point (Zl, Z2, ... , Zp+l) in 8 2p+l will be an element of a unique circle of the form (2.88)
.;~
":J.
:
(zdzj, ...,zjdzj,ZHdzj"",zp+dzj)
(2.90)
Horizontal Geodesic
This coordinate system maps the open set Uj onto CP ~ R 2P. Patching these charts together makes CpP into a differential manifold. We can summarize this as follows: Vertical Orbits
Definition 2.2.10. We define complex projective pspace, denoted by CpP, to be the set of complex lines through the origin in C p+ 1 as in formula (2.86) above. This space can be naturally identified with the set of great circles of S2p+1 defined by formula (2.88).
FIGURE 2.7. Complex projective space represented as a space of circles on the
It remains to construct a metric on the manifold CPp. Rather than beginning at the local level, so to speak, with the construction of the metric tensor, it will be more convenient to define geodesic distance globally on CpP and to note that it leads to a Riemannian geometry on the differential manifold. Let us contract our notation a bit more here by letting z stand for the full vector (Zl, ..., zp+d, which lies in S2p+l. Similarly O(z) will be the element of CpP in which Z lies. Now suppose we wish to define the geodesic distance between two elements O(z) and O(w) of CPp. Write w = (w1, ...,wp+d. We could naturally define the distance between O(z). and O(w) to be
The inner product we are working with here is the inner product on S2P+l as embedded in R 2 p+2 . We can write this in terms of the Hermitian inner product on S2p+l as embedded in Cp+1. This becomes
d[O(z),O(w)] = inf [d(x,y)] : x E O(z);y E O(w)]
(2.91)
where d(x, y) is the geodesic distance on S2p+1 from x to y. We can intuitively think of this formula as saying that the distance from one great circle to another is the shortest gap between them. See Figure 2.7. Now, while this is a perfectly welldefined quantity, there is no reason a priori to suppose that this satisfies the properties that a distance measure, or metric, has. In particular, the triangle inequality has to be checked carefully. The triangle inequality does hold, in a sense, because of the symmetry of the sphere S2p+1. The minimum can be achieved at every value of z by minimizing over y , or correspondingly, at every value of y by minimizing over x. The reader should note the similarity between our construction here and the Procrustean minimization of formula (1.18) in Section 1.3 of the previous chapter. The differences in notation and context should not disguise the fact that the geometric situations are equivalent. In Section 1.2, the points on the sphere were preshapes and the orbits or great circles were shapes. We proceed similarly. Writing the geodesic distance on S2p+1 explicitly, we have
d[O(z),O(W)\
inf [cos 1 « z , y >)
x E CJ(z),y E O(w)]
(2.92)
sphere. In this picture, a small portion of a sphere is seen with circular arcs (the orbits) displayed vertically. Prom a given point on the left orbit of the picture, a variety of geodesics (great circle arcs) can be drawn to the right orbit. A "horizontal" geodesic will have the shortest length and will meet "vertical" orbits at right angles. The distance between two orbits is the shortest great circle path from one arc to another.
< x, y >
~ !Il( 4: x, y:») ~ !Il (~X;Y;)
(2.93)
where Xj and Yj are the jth complex coordinates of x and Y respectively. The next observation we make is that the minimization can be achieved by fixing x = Z and writing Yj = ei 8wj, where i = y"=1, minimizing over 0:::; 0 < 21f. Thus
d [O(z), O(w)]
~;nr {ooo. [!Il % Z;V"wj)] ,0 < 0 < 2~}
(2.94)
We can perform the minimization by maximizing the sum with respect to O. Now R [eW(zjwj)] = cos(O)R(zjwj) + sin(O)8'(zjwj) (2.95) and so the maximum can be found by differentiating with respect to 0 and setting the result equal to zero. This yields ,",p+l
"
L.Jj=l ZjWj
+1 L.Jj=l ZjWj"I l ,",p
(2.96)
Plugging this in, we see that (2.97)
62
2.2 Differential Geometry
2. Background Concepts and Definitions
This is the famous FubiniStudy metric on CPp. As inSection 1.3, when considering distances between shapes, we note that the maximum distance between elements of CpP is IT /2. In addition, the righthand side in this distance formula does not depend upon the specific choice of z and w within the orbits O(z) and O(w). The modulus operation nullifies the effect of this selection, which corresponds to multiplication of the coordinates by a common complex factor of modulus one. As this provides us with a metric on CpP we can now consider the geodesics on this manifold. In Section 2.2.15, we found that the geodesics on RpP were images under the covering map A of geodesic great circle paths of SP. It is natural to consider whether this is the case here. In fact, the geodesics of GpP are images of geodesics on S2p+l. However, they are images of particular geodesics called horizontal geodesics. Intuitively, we think of the orbits of S2p+l as arranged vertically with the mapping S2p+l ~ Cpp as mapping downwards. Thus the horizontal geodesics are always perpendicular to the orbits. See Figure 2.7. These geodesics are great circle paths of S2p+l with the property that they intersect the orbits O(z) orthogonally. More precisely, we can say that a great circle path z(t) is horizontal if for every t the tangent vector z(t) is orthogonal to the vectors of the tangent space of O[z(t)]. It is not the case, in general, that any two points in S2p+l can be joined by a horizontal geodesic. However, if z and ware chosen from O(z) and O(w), respectively, so as to minimize the geodesic distance as above, then z and w can be joined by a horizontal geodesic. The construction of horizontal geodesics will play an important role in Chapter 3, where we shall consider them in greater detail.
2.2.17 Example: Hyperbolic Half Spaces Consider the Riemannian manifold consisting of the upper half space in RP given by (2.98) and metric tensor gjj(Xl' ... , xp) = X;2 (2.99) for all j = 1, ... ,p and gjk = 0 for all 1 S; j f k S; p. The reader will notice the similarity between this space and ordinary Euclidean space. The major difference is the appearance of the last coordinate in the denominator of the diagonal terms of the metric tensor.
w
FIGURE 2.8. The Poincare Plane. Geodesic paths in the Poincare Plane are the arcs of circles that meet the xaxis at right angles. In the limiting form, as the radius goes to infinity, these circles become vertical lines, which are also geodesics.
the EulerLagrange equations, we find that the geodesic paths of HSP are half circles orIines that meet the boundary x p = 0 orthogonally. An important special case is p = 2, which is called the Poincare Plane. See Figure 2.8. It is convenient to represent HS 2 using complex coordinates as HS 2 = { z E C : !2S(z»O}
with
ds 2 =
The family of hyperbolic half spaces HSP represents the negative curvature counterpart of the family of positively curved spheres SP. Solving
Idzl
(2.100)
2
(2.101)
[~(z')J2
Using this complex notation, we can calculate the geodesic distance between two points z and w in HS 2 by integrating ds, given by formula (2.101), along a geodesic path from z to w. As we noted above, these geodesics are circles that are orthogonal to the real axis (with vertical straight lines as the limiting case). Let z and w lie on a geodesic circle centered at a with radius r. As the circle is orthogonal to the real axis, the point a must be a real number. Let rays be drawn from a to z and w making counterclockwise angles {3z and {3w with the real axis. A simple calculation will show that the geodesic distance from z to w is a function of {3z and (3w alone, the quantities a and r disappearing from the final answer. To see this, note that we can write
z = a
+
r cos({3z)
+
i r sin(,Bz)
(2.102)
w = a
+
r cos({3w.)
+
i r sin(,Bw)
(2.103)
and Definition 2.2.11. The space HSP with the metric tensor of formula (2.99) is called the hyperbolic half space of dimension p.
63
where i = yCI. Then the geodesic distance from z to w is given by
l
w
z ds
rPw
= Jp:
csc({3) d{3
=
{ [1 log
cos({3w)] sin({3z) } [1  cos({3z)] sin({3w) .
(2.104)
64
2,
2.2
Background Concepts and Definitions
where we choose the direction of integration so that 0::::; f3z < f3w ::::; rr. From formula (2.104), we can see that the real axis C;S(z) = 0 is not really a boundary at all, but rather an infinite horizon. Half circles that are geodesics in the upper half plane HS 2 have finite length as measured by Euclidean geometry, but have infinite length when measured using the hyperbolic formula of (2.104). The difference is a consequence of the appearance of C;S(z) in the denominator of the formula for the metric tensor. This has the effect of greatly inflating distances compared to the Euclidean metric between points close to the real axis. As we noted above, the geodesic curves of HS 2 include not only the circles of the half plane that meet the real axis orthogonally, but also vertical lines of the form W( z) = constant. These can be thought of as geodesic circles that have infinite radius. For points z and w connected by such a geodesic, formula (2.46) must be interpreted with some care. As ~(z) = ~(w), we can simply integrate formula (2.101) along the imaginary coordinates on which they differ. Alternatively, we can take the limiting form of formula (2.104). In either case, we find that the geodesic distance from z to w is equal to
l
z
w
ds =
l\.'f(W) \.'f(z)
du

[C;S( W)] C;S(z)
log  
=
U
(2.105)
where C;S(w);::: ~(z) > O. This particular formula will have an important role to play when we examine shape variation due to affine transformations in Chapter 3. The transformation zi z + i  (2.106)
z+i
Iwi < 1}
(2.107)
arcs of circles that meet the boundary of the disk at right angles. In the limiting form, as the radius of these arcs goes to infinity, they become diameters of the disk and are also geodesics.
this surface and HS 2 so that the geodesics of HS 2 correspond to the geodesics of the curved surface. To do this, we can construct the Poincare Trumpet HT 2 . It is convenient to use real coordinates (Xl, X2) on the upper half plane of R 2 in this case. For arbitrary 10 > 0, define the functions h,h:R+R (2.109) by 10
hex) = 
(2.110)
X
and
hex)
=
Jx 2 X
10 2
+ log
(x
2 + y'x .
10 2)
(2.111)
Now define
See Figure 2.9. This mapping defines an isometry between HS 2 and HD 2 when the disk is endowed with the metric (2.108)
65
FIGURE 2.9. The Poincare Disk. Geodesic paths in the Poincare Disk are the
maps the points of the Poincare Plane HS 2 onto the the Poincare Disk
HD 2 = {w E C :
Differential Geometry
(2.112) Formula (2.112) maps the region of the Poincare Plane where err eJr to a surface
< Xl
::::;
(2.113) This formula can be derived in a straightforward manner by doing a change of variables from z to w on formula (2.101). It can be checked that the real axis of the Poincare Plane is mapped to the circle Iwl = 1 on the Poincare Disk. This circle becomes its circle at infinity. In addition, the geodesic half circles and lines of HS 2 are mapped to circles and lines in HD 2 that are orthogonal to the circle Iwl = 1. As an additional way of representing the geometry of HS 2 , we might wish to construct a curved surface in R 3 and a correspondence between
in R 3 . See Figure 2.10. While this representation is perhaps the most intuitive way to represent a space of constant negative curvature, the Poincare Trumpet is the least satisfactory in other respects. If the representation is extended to the entire half plane then the mapping ceases to be 11. The mapping of the entire half plane onto the trumpet is, in fact, a covering map that wraps the half plane infinitely many times around the trumpet. Thus the correspondence is only locally correct.
66
2.4 Problems
2.. Background Concepts and Definitions
67
3. An interesting surface called the Moebius strip can be embedded in the interior of the doughnut T from Problem 2 above. LetM 2 be the set of all (Xl, X2, X3) such that
(r  2)2 + x~ ~ 1
X3
sin(9/2) =(r  2) cos(9/2)
(2.115)
where (r,9) are the polar coordinates of (Xl, X2). This is, in fact, a manifold with boundary. The manifold proper is constructed with strict inequality above. Show that the boundary of M 2 is diffeomorphic to .81 . (If we glue the boundaries of two separate copies of a Moebius strip together we also get a manifold without boundary. This manifold is called the Klein bottle K 2 . ) 4. Following from Problem 3 above, we note that another manifold with boundary whose boundary is 8 1 is the disk D 2. This is the set of all (Xl,X2) such that + x~ .~ 1. As the boundary of D2 is diffeomorphic to the boundary of M 2 from Problem 3 above, in principle (given four dimensions to do it in), we could glue the boundaries together by fusing diffeomorphic points. If the two surfaces were cut out from paper we could try to tape their boundaries together. However, as we progressed with the taping in three dimensions we would simply run out of room to do it in. In four dimensions there is enough room. Show that the resulting manifold without boundary is diffeomorphic to the projective plane RP2.
xI
FIGURE 2.10. Hyperbolic geometry representation in three dimensions: the Poincare Trumpet.
2.3 Notes The reader looking for a good introduction to differential geometry may be somewhat overwhelmed by the variety of books that are formal introductions to the subject but make few concessions to the reader who is not trained in abstract mathematics. Such a reader would be well served by looking at the book by Guillemin and Pollack [78] and the book by Morgan [122]. For a general overall introduction to differential geometry, see Spivak [163].
5. Show that the geodesic paths on the sphere 8 2 are arcs of great circles found by slicing the sphere with a plane through the center of the sphere. 6. Consider the cylindrical surface in R 3 defined as the set of all (Xl, X2, X3) such that + = 1 with 00 < X3 < +00..This surface is also represented as 8 1 x R. Show that the geodesics of 8 1 x R are helixes of the form
xI 4
Xl(t)
2.4 Problems
= cos(at)
X2(t) = sin (at)
(2.116)
for arbitrary real values a and b. 1. The Hairy Ball Theorem says that for any continuous tangent vector field on a sphere 8 2 there is some point on the sphere at which the vector field vanishes. Is the analogous result true for the torus 8 1 x s:
2. We can construct a twodimensional surface that is diffeomorphic to the torus 8 1 x 8 1 as follows: Let T be the set of all points (Xl, X2, X3) E R 3 such that
(JXI +x~ 
2)2
+ x~
= 1
7. Prove that formulas (2.29) and (2.30) make tangent vector summation and scalar multiplication well defined. That is, show that the equivalence classes of paths defined for x(to) +z(to) and X x(to) do not depend upon the coordinate system used. Furthermore, show that if y(to) = x(to) and w(to) = z(to) then as defined by (2.28) and (2.29) we have
x(to)
(2.114)
This is the standard doughnut shape. Show that T is diffeomorphic to 8 1 x 8 1.
+
z(to)
y(to) + w(to)
(2.117)
and (2.118)
3 Shape Spaces
3.1
The Sphere of Triangle Shapes
In this and the next two sections, we shall develop a geometric theory of shape due to Kendall [90]. Consider three landmarks (3.1) in the complex plane such that at least two of the three landmarks are distinct. We shall now consider how to naturally represent the shape of the triangle with vertices at Xl, x2, and X3. It can easily be seen that the shape of the triangle can be represented as the complex number Z
=
2X3 
(Xl
X2 
+ X2) Xl
(3.2)
provided that X2 f Xl' The point z in the complex plane has the following interpretation. The triangle XlX2X3 has the same shape as the triangle whose vertices lie at the three points 1, +1, and z. Thus to encode the shape of the triangle we need only move two points, say Xl and X2, to standard positions using a similarity transformation and record the position of the third point under this transformation. The real and imaginary coordinates of z, which determine the shape of the triangle, are called Bookstein coordinates, after F. Bookstein [19], who popularized them. See Figure 3.1.
70
3.1 The Sphere of Triangle Shapes
3. Shape Spaces
71
c real axis
FIGURE 3.2. The stereographic projection. On the complex plane C a sphere sits so that its south pole is at the 'origin. For any point z E C we draw the line passing from the north pole of the sphere through the surface of the sphere at w and meeting the complex plane at z, The stereographic projection thereby puts the points of the closed plane and the sphere into 11 correspondence by mapping z to w. The stereographic projection maps the point at infinity in the complex plane to the north pole of the sphere.
1
+1
FIGURE 3.1. Bookstein coordinates for three planar points. A triangle of landmarks Xl, X2, and Xa is translated, rotated, and rescaled so that the base points Xl and X2 are mapped to 1 and +1, respectively, in the complex plane. The third point xa is then mapped to a point z that encodes the shape information in the triangle. The real and imaginary parts of z are called the Bookstein coordinates.
Such a coordinatization is not without its deficiencies, and it is these that we shall now consider. The most obvious difficulty in using the coordinates of z is that the representation breaks down if Xl = X2. Note that if Xl = X2 and X3 is a distinct point, then the shape of the triangle is perfectly well defined even if the Bookstein coordinates are not. A related problem is that the use of Xl and X2 to standardize a side of the triangle is rather arbitrary. One of the other pairs of points could just as well be chosen. Now suppose Xl = X2 and that X3 is distinct from the other two points. Then the shape of this triangle is most naturally interpreted as z = 00, the point at infinity in the complex plane. So the representation of shape is nondegenerate for all shapes provided this point' is included. The complex plane, with the point at infinity added, is topologically equivalent to a sphere. Putting this another way, we could say that if a point is removed from the sphere S2 the resulting set is homeomorphic to the plane. The complex plane, together with its point at infinity, is called the closed complex plane. A standard mathematical tool that puts the closed complex plane into 11 correspondence with the points of a sphere is the stereographic projection. See Figure 3.2. As the point z in Figure 3.2 follows the locus of a circle in the complex
72
3.1 The Sphere of Triangle Shapes
3. Shape Spaces
plane, the point W on the sphere also follows the locus of a circle, although not necessarily a great circle. As lines in the closed complex plane can be regarded as circles of infinite radius passing through 00, we find that as z follows the locus of a line in the plane, the corresponding point w follows the locus of a circle passing through the north pole. The class of circles generated by such loci for w is the full class of all circles on the sphere. An interesting class of transformations emerges when we look at rotations of the sphere. Suppose we rotate the sphere so that w goes to some point w'. Correspondingly, the point z will move to some point z' elsewhere in the closed complex plane. The transformation z t z' is an example of a type of transformation called a Moebius transformation or a linear fractional transformation, whose general form is z
t
az C Z
+ +
b d
points are antipodal. We can also look for clues to the role of the stereographic projection. Note that the Procrustean distance given in formula (1.21) is indifferent as to the labeling of the landmarks Xl, X2, and X3, provided all triangles are relabeled consistently. For example, using n = 3 in formula (1.21), we could interchange Tn and TIZ and the distance d(aI, az) would not change provided that we similarly interchanged T2l and TZZ. Another way of saying this is that the group of relabelings of landmarks is an isometry of the shape space ~2'. So let us consider how the group of relabelings of XIXZX3 induces transformations on Bookstein coordinates for triangle shapes. If we switch Xl and Xz then z, as defined by formula (3.2), is mapped to z. This is an isometry of the complex plane. However, if we switch Xl and X3 then the point z is mapped by the transformation
(3.3)
where a, b, c, and d are complex numbers such that ad i= be. Just as the class of rotations of the sphere maps circles to circles, so the Moebius transformations map circles in the plane, including straight lines as circles of infinite radius, to circles or straight lines. Using the stereographic projection, we can represent the shape of any triangle XIX2X3 as a point on a sphere in a topologically natural way. For any such triangle, we compute the point z given by formula (3.2) whose real and imaginary parts are the Bookstein coordinates of the shape. We then map z to a point w on the sphere by a stereographic projection. This takes us partway towards the goal stated in Chapter 1, namely the representation of shapes as points on manifolds. However, we are not yet finished. Using the stereographic projection we can make a strong case for the argument that the space of triangle shapes should be homeomorphic to a sphere. However, topological considerations can tell us nothing about distances between shapes. In order to construct a satisfactory representation of triangle shapes as points on spheres we need to find a representation of triangle shapes such that the Procrustean metric of formula (1.21) in Chapter 1 is equivalent to geodesic distance on a sphere. At this stage we have no guarantee that this can be done, and even less of a guarantee that the stereographic projection will be instrumental in the construction. If there is a representation on a sphere that works, we can easily see what the radius of the sphere must be. In Chapter 1, we found that the maximum Procrustean distance between any two triangle shapes was 1f /2. If this is interpreted as a geodesic distance on a sphere, then the radius of the sphere wou.ld be equal to 1/2, and such shapes would be antipodal points on the sphere. In fact, we can find two such shapes. Two triangles whose Bookstein coordinates are z = ±V3i are equilateral triangles that are reflections of each other and have a Procrustean distance of tt /2 from each other. Thus we seek a representation on a sphere in which these two
73
Z
t
+ 3 z  1
z
(3.4)
which is an example of a Moebius transformation of the complex plane. In a similar way to the above, if we switch the triangle points Xz and X3 then the induced transformation of shape becomes
z
3z
(3.5)
t
1+z
which is also a Moebius transformation. In fact, our first transformation z t z is also a special case of a Moebius transformation. The group of relabelings of shapes is the set of six transformations of the complex plane that can be written as the arbitrary composition of these three Moebius transformations. It is no coincidence that the Moebius transformations of the complex plane arise in relabeling triangle landmarks and also as the images under stereographic projection of rotations of the sphere. In both cases we are dealing with isometries  in the former case the isometries of ~~ and in the latter case isometries of the sphere. The type of transformation that we are seeking should be a stereographic projection from the closed complex plane onto a sphere of radius 1/2 taking the two equilateral triangles into antipodal points! Suppose the shape of triangle XIXZX3 is displayed by Bookstein coordinates as a point z in the closed complex plane. Now define
WI
1 lzl z/ 3 = 2(1 + Izl z /3)'
~(z)/V3
R(z)/V3
Wz
=
1 + Izl z /3'
W3
=
1 + Izl z /3
(3.6)
Then (3.7)
is a stereographic projection of the triangle shape in Bookstein coordinates onto a sphere of radius 1/2 centered at the origin in R 3 . The mapping
74
3.
3.1
Shape Spaces
1
3
2
• • •
The Sphere of Triangle Shapes
13
••
75
2
•
great circle of
collinear triangles
1
32
•
••
3
2
• • •
• • •
••
•
3
12
3
1 FIGURE 3.3. Spherical geometry for three planar points. The shape of any triangle X1X2X3 is encoded in Bookstein coordinates z as a point in the closed complex plane and then mapped by a particular stereographic projection to the sphere. There are two antipodal points on the sphere that correspond to the two equilateral triangles of landmarks in the plane. Passing through these two antipodal points are three great circles that correspond to the isosceles triangles of landmarks  each great circle characterized by the choice of vertex at which the isosceles angle occurs. A family of isosceles triangles around one such great circle is displayed around the outside ofthe sphere. Triangles of aligned landmarks (i.e., collinear triangles) are. to be found on the great circle of the sphere that is equidistant from the ~wo equilateral triangles and orthogonal to the great circles of isosceles triangles.
1
2
FIGURE 3.4. A closeup look at the great circle of collinear triangles from Figure 3.3. Three points separated by arcs of 120 0 mark the collinear triangles where two landmarks are coincident. Halfway between these points are the collinear triangles of equally spaced points.
76
3.
3.2 Complex Projective Spaces of Shapes
Shape Spaces
from the preshape of
XlX2X3
to the shape of the triangle is a mapping
(3.8)
It is helpful to study the sphere S2(1/2) by finding the coordinates of interesting triangles on it. For example, there are two equilateral triangles represented by antipodal points on the sphere at W3 = ±(1/2). That there are two equilateral triangles rather than one is a consequence of the fact that triangle shapes are not identified with their reflections. Halfway between these antipodal points are the shapes corresponding to W3 = O. These shapes lie on a great circle of S2(1/2) that is the set of collinear triangles. In other words, these are the triangles that have a straight angle at one of the vertices. Included in this set are the three shapes corresponding to triangles where two of the points are coincident and the third point is distinct. These shape points are equally spaced at angles of 27f/3 radians around the great circle W3 = O. See Figure 3.4. The reader should make a careful comparison of Figures 1.3 and 3.4. In both figures, we see collinear triangles of landmarks displayed as points around a circle. However, there is an important difference. In Figure 1.3, the preshapes of triangles that are reflections of each other are distinct antipodal points of the circle. However, as we argued earlier in Example 2.2.15, the shapes of collinear triangles in the plane lie naturally on a real projective space and not a sphere. As it happens, the real projective space Rpl is isometric to the circle Sl(1/2). So each pair of antipodal points of Figure 1.3 is represented as a single point in Figure 3.4. For example, the earlier figure has six points that represent the shapes of triangles where one landmark is at the midpoint between the other two. However, Figure 3.4 has only three such points around the circle. The preshape space of Figure 1.3 is the unit circle s', whereas the space of collinear shapes in Figure 3.4 is the real projective space Rpl ~ s (1/2). . To obtain the great circle distance on S2(1/2) between any two shapes, we use the inner product between vectors on S2(1/2). If u = (Ul,U2,U3) and v = (VI, V2,Va) are two points on S2(1/2) then the geodesic distance from u to V is given by formula (2.69) using r = 1/2. We obtain d(u,v) =
21
cosl(4 < u,v
»
(3.9)
We leave the reader to check that the great circle distance defined by this formula is equivalent to that of Chapter 1. See Problem 2. Before we turn to the study of 2;2 it is worth considering some of the geometry of the sphere and its relationship to Bookstein coordinates. The geodesic paths are the shortest paths between points. As we noted, these are the arcs of great circles on S2(1/2). To find the corresponding paths in Bookstein coordinates, we need to construct the images of the great circles under stereographic projection. Any circle on the sphere S2(1/2}
77
is mapped by the inverse of the stereographic projection defined by (3.6) to a circle or a straight line in the plane. Among these, the images of the great circles are a subset. The zaxis of collinear shapes is an example of a geodesic in Bookstein coordinates. To find the others, note that any two great circles of S2(1/2) will intersect in antipodal points. In fact, we can characterize a great circle of the sphere as a circle meeting the equator of collinear shapes in antipodal points. Now in Bookstein coordinates, two points Zl, Z2 E e are images of antipodal points on the sphere if Z2 = 3/ zj. This can be checked by plugging z = 3/ zj and Z = Zl into the coordinates of the stereographic projection in formula (3.6). After some rearranging, we see that the resulting stereographic coordinates become the negatives of each other. Therefore, any circle in the plane of Bookstein coordinates that passes through points of the form a and 3/a on the real axis will be the stereographic image of a great circle of S2(1/2).
3.2
Complex Projective Spaces of Shapes
In this section we shall study the spaces 2;2 where n ::::: 3. As we shall see, the sphere of triangle shapes described in the previous section is a special case of a complex projective space having two real dimensions. We will continue to identify landmarks Xj in the plane with elements of the complex plane e. Suppose (Xl, X2, ... , X n ) are n such landmarks, at most n  1 of that are coincident. To discover the information in this configuration of landmarks that is invariant under Sim(2), we first remove the effect of translations by centering the points about their centroid x yielding' (Xl  ii ; X2  x, ..., X n  x) (3.10)
en
This vector lies in a subspace of having n  1 complex dimensions or 2n  2 real dimensions. The effect of multiplication of these variables by a complex nonzero quantity >. i
!
[>'(Xl  x), >'(X2  x), ..., >'(xn

x)] .
(3.11)
is to scale the centered points by 1>'1 and rotate them by arg(>.). To remove the effect of complex multiplication, we identify all such multiples and decl~re them to lie in the same equivalence class. shape space 2;2 can be identified with the set of complex lines So, througj the origin in the subspace
tIl
!
Fn  l
=
{(Xl, ...,
X
n) E en : t X j = O}
(3.12)
j=l
which has n  1 complex dimensions. This looks very similar to complex projective space Cpn2, as given in Definition 2.2.10. The difference is
78
3.3
3. ShapeSpaces
that we are considering complex lines through the subspace Fn,l rather than en. However, this difference turns out to be superficial, because we can construct a linear isometry from Fnl .to e n l that maps complex lines through the origin in the subspace Fnl to complex lines through the origin in enl. To construct this linear isometry, we define (3.13) for 1
~
j
~
n  1. The mapping
Landmarks in Three and Higher Dimensions
79
The change of radius is a secondary consideration here. A continuous function from S3 to S2 of this kind is an example of what is known as a Hop! fibration between the spheres. In general, it is impossible to find continuous mappings from a sphere of one dimension to a sphere of a lower dimension that are locally projections of this kind. However, there are special dimensions for which it is possible. From three dimensions to two dimensions is one such case. Such limitations in. dimensions already give us a clue that the Procrustean approach to the shape of a general number of points in general dimensions will not be as smooth a theory as for the shapes of points in dimension 2.
(3.14) is a linear isometry from F n l to e n l that preserves the complex lines of (3.11) above. Under the identification established by (3.14), we can see that the definition of the Procrustean metric in Section 1.3 is completely parallel to the definition of the FubiniStudy metric in Section 2.2. In particular, formulas (2.92) and (1.18) yield equivalent metrics under the identification of (3.14). Thus we have proved the following result: Proposition 3.2.1. The shape space ~2 endowed with the Procrustean metric is isometric to the complex projective space ep n  2 .
So, the shape spaces ~2 are Riemannian manifolds such that geodesic distance between points in the shape space is equivalent to the Procrustean metric defined in Chapter 1. A technical note on this point is that the Gaussian curvature of ~2' is a constant throughout the manifold and equal to 4. By contrast, the sphere s2n3 of preshapes has a constant positive curvature equal to 1. (The reader who is not familiar with Gaussian curvature on manifolds should rest assured that this notion will not playa large role in our exposition of shape geometry.) A special case of this was seen previously for the sphere of shapes, which is required to have radius 1/2. In general, the mapping (3.15) of each preshape into its shape equivalence class becomes what is known as a Riemannian submersion, a local projection that will be described in greater detail in the next section where we shall consider the general spaces ~;.
A special case of our construction is quite famous in differential geometry. We have seen that ~~ is isometric to the sphere S2(1/2). Thus the mapping from each preshape to its corresponding shape is equivalent to a mapping (3.16)
3.3
Landmarks in Three and Higher Dimensions
3.3.1
Introduction
So far, we have only considered the shapes of landmarks in two dimensions. However, the shapes ofsolid objects are of common interest, and are most naturally represented by landmarks in three dimensions. Landmarks in four and higher dimensions are of interest in multivariate statistics, where the shapes of multivariate data sets provide information about normality, linearity, and correlation between variables. Let Xl, X2, ... , X n be n ~ 3 landmarks in RP, where p ~ 3. We shall suppose that at least two of these landmarks are distinct, so that L Ilxj  xl1 2 > O. The standardization of the location and scale of these n landmarks can proceed in a manner similar to the twodimensional case. The preshape T . can be constructed by centering the landmarks about their centroid x and by rescaling the centered configuration of landmarks so that 2:: Ilxj  xl1 2 = 1. Thus the preshape T of Xl, ..., X n can be seen to be an element of the sphere S np* p.' , l = {( Yl, ... , Yn ) E I
Rnp : L.J ~ Yj =
2 0, ~ L.J II Yj 11 = 1}
(3.17)
The space ~; of shapes of Xl, ... ,xncan now be formally identified with t~k collection of equ.ivalence classes in S~Ppl of all preshapes sharing' a common shape. For any preshape TE S~PPl, let OCT) be be the/set of all preshapes T' that have the same shape as T. For example, in dimension p= 3 the special orthogonal group SO(3) is simply! the group of rotations about the origin in threedimensional space. Let h ESO(3). Suppose that the landmarks Xl, ..., X n have preshape T. For j = 1, ... ,n, let xj = h(xj) be the jth landmark rotated by h. So the landmarks x~, ..., x~ are a rotated version of Xl,'''' X n. If T' is the preshape of X~ , ... , x~, then T' will be an element of OCT). The converse will also follow. If X~, ... ,x~ have a preshape T'E OCT), then there will exist a rotation u e SO(3) such that xj= h(xj) for all j.
80
3.3 Landmarks in Three and Higher Dimensions
3. Shape Spaces
Such equivalence classes can be defined similarly in higher dimensions p using the special orthogonal group SO(p). Then we can define (3.18) We can also define the function (3.19) taking each preshape T to its corresponding equivalence class, or shape, O(T). Now any set of n landmarks in RP can be identified with an element of RnP. Since the elements of the group SO(p) transform the landmarks individually, we can regard SO(p) as a group of transformations on Rnp. Each hESO(P) mapsapoint (Xl,""X n ) in Rnp=(RP)n by the rule (3.20) Interpreted in this way, the group SO(p) becomes a subgroup of the group of special orthogonal transformations on Rnp, namely SO(np). The next thing; to note is that S~pPl is a subset of Rnp, and that transforming according to (3.20), the transformations h E SO(p) map S~ppl onto itself. So SO(p) is a class of isometries of the sphere S~ppl. Moreover, we can write O(T) = {h(T) ; hE SO(p)}
(3.21)
We can introduce a Procrustean metric between shapes in 'E; in a manner similar to the twadimensional case. So for any shapes 0'1 = O(Tl) and 0'2 = 0(T2) in 'E; we can set the Procrustean metric d(0'1,0'2) to be inf{cos1«Tl,T2» : O'j=O(Tj) forj=I,2} (3.22) This is equivalen.t to the definition given in formula (1.18) with the appropriate change in dimension. However, as we shall see, appearances are deceiving here. The extension of the geometry of 'E2' to higherdimensional settings is not as routine as this formula would suggest. One algebraic advantage is lost in the generalization: the algebra of the complex plane is not available for representing the Procrustean metric when landmarks are chosen from three or higher dimensions. The metric of (3.22) does not in itself provide much immediate insight into the topological and differential structure of 'E; .We can construct the topology directly on 'E; without direct reference to the metric d. A subset U of 'E; will be open if and only if sp,;(U) is an open subset of S~ppl. With this topology, the function spn becomes continuous. It follows immediately from this that all the shape spaces 'E; are compact, np p l because they are continuous images of the compact sp h eres S * .
81
Now let us consider the space 'E~ for n = 3,4, .... The Euclidean space R n 1 can be canonically embedded in R" so as to be an (n  1)dimensional subspace of R". This embedding induces a mapping from 'E~l to 'E~ that takes the shape of a set of n points in Rnl c R" into the shape of the same set of points considered as lying in R n . In the case where n = 3 we can see what this does. See Figure 3.5. The shapes of point configurations in R 2 that are relections of each other through some line are not generally of the same shape. However, in R3 a plane can be reflected about some line by a rotation. Thus configurations that are mirror images in R 2 have the same shape when embedded in R3. The shape space 'E~ is the sphere S2(1/2), and the associated mapping into 'E~ identifies every triangle shape with its mirror image in R 2 . In the coordinate notation of formula (3.6) this identifies points of the form (WI,W2, W3) with (WI,W2, W3). Thus 'E~ is topologically a hemisphere with the collinear shapes forming its boundary. This example points out a major obstacle to the study of the shape spaces in general dimensions. On the boundary, the hemisphere is not locally homeomorphic to R 2 as it is in its interior. So a hemisphere is not a topological manifold at all, but must be classified as a manifold with boundary. Generally, the spaces 'E; will have boundaries whenever p 2: n. Even when is a topological manifold, it need not have a natural definition as a differential manifold. Singularities in the smoothness can arise much as one can introduce a crease into a surface. In a private communication to David Kendall, A. J. Casson proved that the shape spaces 'E~+l are all topologically spheres for n 2: 2. That this is the case for n = 2 we have already seen. However, that it should be true for the topology of 'E~+l for n 2: 3 is interesting because it is kno~n that these spaces are not diffeomorphic to the usual spheres of equivalent dimension. The presence of singularities in the differential structure is enough to ensure this. In honor of Casson's discovery, D.G. Kendall proposed that the shape spaces 'E~+l be called Casson spheres. Unfortunately, Casson's proof is not available in the literature although another p~bof has been published. See Le [104J. Le's proof makes use of the Riemannian geometry off the singularity sets of the Casson spheres to I prove Casson's result. See also Carne [38] for an analysis of the geometry of these shape spaces. Let us Ip.ow consider the differential geometry of the general shape spaces 'E;. In oirer to do this we shall need to define the concept of submersion between differential manifolds. We have the following definition:
E;
\ Definition 3.3.1. Let h: MP > N? be a differentiable mapping onto the manifold N", where q:S; p. We say that h is a submersion at a point x E MP when the linear mapping (3.23)
82
3.3 Landmarks in Three and Higher Dimensions
3. Shape Spaces
83
is of Jull rank q or equivalently, when (Vh)x is onto. The mapping h is said to be a submersion provided that it is a submersion at all points xEMP. We have already encountered a number of examples of submersions. For example, the class of submersions includes linear projections Rq
• •
•
• •
•
x
RPq > Rq
(3.24)
mapping (Xl, ••• , X q , ... , X p )
>
(X1, ...,:1;q)
(3.25)
A submersion between manifolds can be regarded as a differentiable mapping that is locally equivalent to a projection. From our point of view, perhaps the most important examples of submersions that we have encountered are the mappings (3.26)
FIGURE 3.5. The effect of embedding a configuration of three planar landmarks into three dimensions. Configurations of landmarks in R 2 which are reflections of each other have different shapes, because transformations that reflect the plane are not elements of the group Sim(2). By contrast, coplanar configurations of landmarks in R 3 that are reflections ofeach other do have the same shape. This is because the group Sim(3) includes 180 rotations of planes in R 3 about an axis in the plane. The shape of three landmarks in R 2 lies in Et while the shape of three landmarks in R 3 lies in Eg. The identification of shapes that are reflections of each other in E~ can be regarded as the identification of points on opposite hemispheres of the sphere of triangle shapes. Topologically, the effect of identifying points in opposite hemispheres is to fold one hemisphere into the other and to glue the two surfaces together. Thus the space of triangle shapes in three dimensions is topologically a hemisphere. 0
taking the preshapes of planar configurations of landmarks to their shapes. In particular, the Hopf fibration from S3 to S2 is a submersion. The problem at hand is to make :E; into a differential manifold in such a way that its atlas is compatible with its topology and so that the mapping spn : S:,pp1 >:E~ becomes a submersion. The detailed conditions under which this is possible are given by Dieudonne [51, Section XVI.10J and will not be explained in detail here. We shall simply note that the submersion can be constructed for some preshapes (and their corresponding shapes) but not for others. The result is that there exists a singularity set within each shape space :E; such that outside this set a local smooth structure can be imposed at all the points, making spn a submersion. The particular locus of this singularity/set within E; is determined by the failure of the group SO(p) to a,et,!reely on the sphere S:,pp1 as defined below. //' Definition ;r:3.2. Let H be a group oj transformations h: MP > MP on a manifold MP. We say that H acts freely onMP iJ the only transjormaiion. h E H for which hex) = X [or some X E MP is the identity transformation. In other words, if H is free, then every transformation h that moves some point of the manifold will move all points of the manifold. For example, the group SO(2) acts freely on s, whereas the group SO(3) does not act freely on S2. As Le and Kendall [105J have noted, the singularities in :E~ arise because they are the images under spn of preshapes at which the action of the group SO(p) on S:,ppl is not free. For example, let p:::: 3 and consider
84
3.
3.3
Shape Spaces
Landmarks in Three and Higher Dimensions
85
a set of n ::::: p+1. points Xl, X2, ••• , X n lying in RP. We center the location of the points by subtracting the centroid x. Now suppose that there exists some (p  2)dimensional subspace in which the centered points Xl 
X,
X2 
X, ... ,
Xn
 X
(3.27)
all lie. Then there exists a special orthogonal transformation of RP that is not the identity transformation and that leaves this (p  2)dimensional subspace fixed. To illustrate this, let us consider what happens in p = 3 .dimensions. Put rather simply, we can say that it is possible to rotate a configuration of landmarks without changing the orientation (i.e, leaving the preshape fixed) provided the landmarks all lie along the axis of rotation. This is in contrast to dimension two, where a configuration cannot be left invariant under a rotation unless all the landmarks are coincident. Suppose that Xl, ... , X n are collinear landmarks in R 3 , and that n::::: 3. Then the centered landmarks Xl x, ... , X n x will all lie along a line passing through the origin. Now suppose a rotation hE 80(3) is chosen that has this line as its axis of rotation. Then h will leave the vector (Xl  X, ... , X n  x) fixed under the transformation (3.28) In addition, the rotation h will induce a transformation on the sphere of preshapes 8~n4, mapping the preshape of Xl,'''' X n to the preshape of h(XI), ..., h(x n ) . If the transformation in (3.28) leaves centered landmarks Xj  x fixed, the same will be true of the preshapes. Thus 80(3) does not act freely on 8~n4. See Figure 3.6 for an illustration of this. In general dimensions the group 80(p) will fail to act freely on 8~ppI when n, p ::::: 3. The singularity set in ~; will be the set of those shapes of landmarks Xl, ... , x n which lie, when recentered as in formula (3.27) above, in a (p _. 2)dimensional subspace. In the fivedimensional Casson sphere ~~, for example, this subspace is onedimensional. Therefore, the singularity set is the subset of collinear shapes.
3.3.2
Riemannian Submersions
Let us now turn to the problem of defining a metric tensor 9 on the open subset of that is the complement of the singularity set. In order to describe this, we have to define a type of submersion, called the Riemannian submersion, which is specific to the theory of Riemannian manifolds. Suppose MP and N", for p > q, are Riemannian manifolds with metric tensors gM and 9N respectively. These metrics define inner products on the tangent spaces T(MP) and T(Nq) respectively. Now let h : MP 4 N" be a submersion. Then for each X E MP and y E N? such
E;
I \ \
FIGURE ~.6. Singularity sets in shape spaces. If a set of landmarks in R 3 is collinear,. in the top diagram, then rot~tio.ns about the line through the landmarks wzll leave the landmarks fixed. This is an example of the failure of the group of ro ations to act freely. Singularities in the shape space Ej occur at points corres~onding to such configurations of landmarks. Singularities in highdimensional mfLnifolds are difficult to understand, although singularities do appear in low enough dimensions to help us visualize them. Two types of singularities . are illustrated in the middle and bottom diagrams. In the middle diagram, we see a topological singularity in a space. The singularity is the set of points where two surfaces intersect. At these points, the space fails to be locally homeomorphic to R 2 , and is not a topological manifold. However, if this intersection set is cut out, the remaining set does become a topological manifold. In the bottom diagram, we see another type of singularity set in a surface. In this case, the singularity is in the smoothness, or differential structure, of the manifold. Unlike the middle diagram, there is no topological singularity. The singularity set in the shape space E~ is of this nature. This shape space is topologically a sphere, but contains a higherdimensional analog of the type of singularity displayed in the bottom diagram. We cannot do differential geometry [i.e., construct tangent vectors or set up a metric tensor) at the singularity set, but we can do it elsewhere.
86
3.4 Principal Coordinate Analysis
3. Shape Spaces
that y = hex) the derivative
1 h [hex)]
vx
(3.29) is a linear transformation of full rank. We shall say that h is a Riemannian submersion if (Dh)x is equivalent to an orthogonal projection for all x E MP. The following more precise definition can be given:
M
87
P
..L
V X
Definition 3.3.3. Let h: MP + Nq be a submersion as described above, and let x E MP. We define the vertical subspace Vx(MP) to be that subset of Tx(MP) defined by (3.30) h
This is the kernel of the mapping (Dhk The vectors of the vertical subspace are called the vertical tangent vectors at x E MP. Orthogonal to the vertical subspace is the horizontal subspace, which we now define.
•
Definition 3.3.4. The horizontal subspace V/(MP) is defined to be the set (3.31)
where the inner product is calculated in Tx(MP) using the metric tensor gM' Similarly, the vectors of the horizontal subspace are called the horizontal tangent vectors at x E MP. It is easy to see that any tangent vector at x can be uniquely written as a vector sum of a horizontal and a tangent vector that are orthogonal to each other with respect to the metric tensor gM' See Figure 3.7 for an illustration of the horizontal and tangent vectors at a point x E MP. Using concepts of horizontal and vertical tangent vectors, it is now possible for us to define the concept of a Riemannian submersion. Definition 3.3.5. A submersion h: MP submersion at x if
+
Nq is said to be a Riemannian
(3.32)
is a linear isometry when these spaces have metric tensors gM and gN respectively. We shall say that h isa Riemannian submersion if h is a Riemannian submersion at all points x E MP.
q
N
hex) FIGURE 3.7. Decomposition of T,,(MP) into vertical and horizontal components.
The basic principle for constructing a metric tensor on 'E~ off the singularity set is to define it so that spn is a Riemannian submersion at all preshapes T E S:,ppl at which spn is a submersion (i.e., preshapes outside, the singularity set). Such a metric tensor is uniquely defined. Thus the determination ofthe metric tensor on'E~ is equivalent to the evaluation of the metric tensor on S:'PPl restricted to the horizontal tangent spaces. D.G. Kendall and H. Le have carried out this program to evaluate the metric tensor on the shape spaces. See [104] and [105]. With this geometry, the geodesics of 'E~ become the images under the mapping spn of the horizontal geodesics of S:'PPl. These are the geodesics x(t) of the sphere for which (3.33) at all points x(t) along the geodesic path.
3.4
Principal Coordinate Analysis
The Procrustean metric and the shape spaces of the previous sections provide very general tools for the representation of the shapes of landmark
88
3. Shape Spaces
configurations as points in manifolds. However, mathematically elegant as these representations are, they represent an impediment to the graphical representation for exploratory data analysis, which much be accomplished in a small number of dimensions. For example, if three landmarks are selected from each of fifty images, then the resulting landmark shapes can be displayed as a configuration of fifty points on an appropriate projection of the sphere E~. However, more detailed descriptions of the shapes will require more lanclmarks from each image, and a correspondingly higherdimensional manifold in which to portray the fifty points. The tools for shape representation that we have been considering can be useful for the exploratory analysis of shapes when they can be coupled with dimension reduction methods that are designed to approximate the highdimensional configuration of the points by low (usually one or two) dimensional configurations whose interpoint distances most appropriately approximate those of the highdimensional configuration. Such methods are called multidimensional scaling. There is considerable reason for optimism about the use of multidimensional scaling, because from formula (1.21) we see that the geodesic distance between two shapes can be quite simple to compute even when the complex projective spaces in which the shapes live are hard to visualize. Suppose Xl, Xz, ..., Xn are elements of some Riemannian manifold MP. We shall let d jk == d(xj, Xk) be the geodesic distance from Xj to Xk. The n x n distance matrix (dj k ) is a symmetric matrix of nonnegative values. (The particular application we have in mind is that where MP is a shape manifold and d is possibly the Procrustean metric given by (1.21).) The task of multidimensional scaling is to find a set of points Xl, X2, ..., Xn E Rq (where usually q = 1 or 2) such that if djk = d(Xj,Xk) then the matrix (djk ) approximates (d jk) in some predetermined sense. The various methods used to approximate (djk) by (djk ) can be used to categorize the types of multidimensional scaling. Broadly speaking, the methods divide into two groups called metric scaling and nonmetric scaling respectively. In metric scaling, the task is to make the distance matrix (djk ) match (d jk) as closely as possible. In nonmetric scaling this requirement is relaxed. A typical criterion is that the distances dj k should be ordered as closely as possible to the ordering of the distances djk . In this section we shall describe a computationally straightforward technique for metric scaling called principal coordina1eanalystnIueToG~wer [74]. This should not be confused with the betterknown term principal component analysis, although the two techniques are related and rely on the common principle of an appropriate eigenvector decomposition of a positive definite matrix. Let us begin with the following problem: Suppose that Xl, Xz, ..., Xn are n points in some pdimensional space that we can take to be Euclidean. The coordinates, or positions, of the points themselves are unknown. How
3.4
Principal Coordinate Analysis
89
ever, the distances djk = d(xj, Xk) between the points are given to us. As the original points are unknown, how can we construct a set of points Xl, X2, ..., Xn, which are not necessarily in p dimensions, with interpoint distances djk = d(xj, Xk), such that djk = djk for all 1::; j, k ::; n? Let us start with any matrix (djk) of interpoint distances. The task of constructing the set of points Xl, ... , xn proceeds as follows:
Step 1. From the distance matrix (djk) we form the association matrix 1 = (1 jk) by defining 1 jk = d]k/2, for all 1::; i, k ::; n. Step 2. In the second step, we standardize the matrix 1 so that its rows and columns sum to zero. This is accomplished by defining (3.34) and
1 ~

(3.35)
1 =  L..J1j
n j=l and then defining the matrix
njk
=
n=
(njk)
by
1jktj  t k + i'
(3.36)
Gower [74] notes the following result, which we state without proof. Proposition 3.4.1. Let (d jk) be a matrix of interpoint distances. Then the matrix n defined in Steps 1 and 2 above is nonnegative definite. That is, the eigenvalues of are nonnegative.
n.
Step 3. In the third step, we construct an n x n matrix whose jth row is the eigenvector Vj corresponding to the jth largest eigenvalue Wj of the matrix n. The eigenvector Vj is standardized so that Wj = Ilvj W· (This is possible because Wj::::: 0 for all i, by Proposition 3.4.1.) We can display this n x n matrix as in (3.37) below.
c==]
Xl
Xz
Xn
W1
Vll
VIZ
V1n
(3.37)
90
3.
3.4 Principal Coordinate Analysis
Shape Spaces
For example, VI = (Vu, Vt2,.." Vl n). At the left of each row, the eigenvalue Wj is listed that corresponds to the eigenvector Vj' Step 4. Reading across the columns of this matrix, we obtain the eigenvectors VI, ..., Vn of the matrix n. However, reading down the rows of the matrix gives us the required vectors Xl, ... ,xn . For example,xl = (Vl1, V2l,..., Vnl)T.
Lemma 3.4.2. The eigenvectors of a symmetric n
~ n, we have
d j k =d(Xj,Xk)
Proof. We can write d;k
= Ilxj 
=
=
n
n
"2: Vrj + "2: Vrk 1=1
(3.41)
djk
xklj2. Expanding this out, we get
n
dJk
To prove this result, we shall need the following lemma, which we state without proof.
..., xn
be the vectors constructed in step
4
and complete the proof. Q.E.D. The dimensionality of Xl, .•., Xn is typically too high for convenient graphical representation. However, the principal coordinate analysis also provides a principal component analysis of these points. The eigenvalues have been ordered in decreasing size from top to bottom in the rows of the matrix (Vjk), thereby ordering the coordinates of Xl, ..., x n from the coordinates along the axis with highest variation (coordinates at the top) to those of lowest variation (at the bottom). So, for example, to choose a twodimensional projection of the vectors Xl,""X n , we can take the 2 x n block consisting of the first two rows of (Vjk) in (3.37). In shape analysis for planar landmarks, we will start with a matrix (d j k ) of interpoint geodesic distances d(aj, ak) between shapes, rather than the matrix of Euclidean distances described above. In this case, d(aj, ak) will be the Procrustean distance between two shapes in 2J~. Now, there is no a priori guarantee that the matrix r! will be nonnegative definite, as in
92
3. Shape Spaces
Proposition 3.4.1, because the interpoint geodesic distances on a manifold satisfy different inequalities from those in Euclidean space. Nevertheless, the matrix n can be calculated from the matrix (dj k ) , and its eigenvalues can be checked. If the Procrustean distances d j k can be approximated by Euclidean interpoint distances, then the largest eigenvalues of n will be positive. So, for example, if the first two principal eigenvalues are positive, then the 2 x n matrix of the first two rows in (3.37) can be constructed. If we define Xj E R 2 to be the jth column of this 2 x n matrix for j = 1, ..., n, then Xl, ..., Xn will be a twodimensional configuration whose interpoint distances approximate (dj k ) . More generally, with k of the eigenvalues positive, we can construct a set of points ii I, . '" xn in R k. (EoLgraphi is positive. Let 0 be the orthogonal n x n matrix whose jth column is the vector ej. Then Wx = 013 x can be shown to be an upper triangular matrix with positive entries down the main diagonal. For the proof of this result, see Problem 7. The matrix n 1 produces an orthogonal transformation of the column vectors of 3", that standardizes the orientation information in 3 x . For this reason, we can call W'" the sizeand shape matrix of the landmarks. Next, we eliminate scale information in Wx by dividing every element of this n x n matrix by the element.In the upper left corner. We need have no fear that this element of the rr\atrix is zero because the elements on the main diagonal of Wx are all pos\tive. Upon dividing every element of Wx by the upper leftmost element, wlj are left with the upper triangular matrix I 1 Z31 Z41 Z(n+l)1
I.··
! Z(n+1)2
o
0
o
Z43
0
0
Z(n+1)3
0
(3.62)
Z(n+1)n
which is the matrix representation of the shape of the landmarks. The reduction to shape coordinates has proceeded via a series of reductions. First, we reduced to the presizeandshape matrix 3 x , then to the sizeandshape matrix Wx, and finally, after standardization, to the shape matrix II x • The reason for the rather strange labeling of the elements of II x is the following. Suppose we define Z1 Z2
= =
(0, 0, 0,
(+1, 0, 0,
, 0)
, 0)
(3.63)
(3.64)
and for 3::; j ::; 11, + 1,
(3.65) Then the simplex: with vertices Xl, X2, , Xn+l (or its mirror image) and the simplex with vertices Zl, Z2, , Zn+t have the same shape. The coordinates defined by (3.63)(3.65) encode the information about the shape of the landmarks Xl, ... , Xn+l' Thus we have the following definition.
101
...%~
(0,0,0)
(1,0,0)
FIGURE 3.12. Generalized Bookstein coordinates for a simplex in three dimensions. The simplex is mapped by a similarity transformation so that the landmarks and X2 are mapped to (0,0,0) and (1,0,0) respectively. The simplex is rotated about the axis through (0,0,0) and (1,0,0) until the third landmark is of the form (Z31,Z32,0) with Z32 > 0. If the coordinate Z43 is negative, the fourth landmark is reflected through the plane of the other three landmarks to make this coordinate positive. Compare this figure with Figure 3.1.
Xl
shall be called generalized Bookstein coordinates of
Xl, ... , X n + 1.
See Figure 3.12. The reader should note that that these coordinates do not generalize Bookstein coordinates in the strict sense because for the case n = 2 the simplex with vertices at Zl, ... , zn+t has its first point anchored at 0 rather than at 1, as was the case in the previous section.
Definition 3.6.3. We define UT(n) to be the set of all upper triangular n x n, matrices II = (II j k ) for which II l l = 1 and for which the diagonal elements II j j are all positive. We shall call the matrix II x of (3.62) the upper triangular shape representation of x = (Xl, ... ,Xn+t), or the UTshape representation of X for short. It is easy to see that UT(n) is closed under matrix multiplication and inversion. Moreover, since the identity matrix is in UT(n) it follows that UT(n) is a group with matrix multiplication. Our next task is to make UT(n) into a Riemannian manifold by constructing a metric tensor on it. Let II x be an element of UT(n). We perturb II x to a neighboring matrix II x+dx' To introduce a metric tensor on UT(n), we need to find the singular values of IIx+dxII;I. The extent to which these singular values differ from each other is a measure of the shape change induced by left multiplication by the matrix IIx+dxII;l. Let
(3.67)
Definition 3.6.2. The coordinates (3.66)
To construct a metric tensor on UT(11,) we need to find an appropriate
102
3.6 Hyperbolic Geometries for Shapes
3. Shape Spaces
quadratic form on the coordinates of dA. There are n eigenvalues of the matrix ATA , and these eigenvalues, as perturbed values of unity, can be written as Aj = 1 + dAj for j = 1, ..., n. Unlike the case for n = 2, these eigenvalues cannot generally be found with simple algebraic expressions. Fortunately this is unnecessary, as the first and second moments of the eigenvalues can be computed from the coefficients of the characteristic polynomial det[>.I  ATA]
= det[>.I  (I + dA T + dA)] = 0
(3.68)
dAjk of the matrix dA. To do this, we return 'tothe characteristic equation given in (3.68) and evaluate the coefficients explicitly. We obtain n
al
we recall that (3.70) and a2 =
L
L
[(1 + 2dA j j) (1 + 2dA kk)  dA~k]
(3.74)
l=5j R n is a diffeomorphism such that M n == 0 and N n == 0 throughout R". Then ti e Sim(n). Proof. As M n == 0 it follows that the Jacobian matrix of h is locally a rescaling of an orthogonal matrix. That is, there exists a positive function A : R n 4 R such that AA is an orthogonal matrix at all points in R ". But because N n == 0 we also observe that Jh is a constant function throughout R". Moreover, Jh == A n . Therefore, A is constant function on R". Thus we can say that h = Aho, where A is a positive scalar, and Ao, the Jacobian matrix of ho , is an orthogonal matrix at all points in n. However, any function h o must be an isometry of R" if Ao is everywhere an orthogonal matrix. To prove this, consider two points x, y E Rn,andconsiderthelinesegment L with endpoints x and y.Let ho(L) be that arc in Rn from ho(x) to ho(y) consisting of the image under h o of all points on the line segment L. Because Ao is an orthogonal matrix, it follows that ho is locally an isometry, so that the length of the path ho(L) is equal to the length of the line segment L. But L has length Ilx  yll, and the length of ho(L) is greater than or equal to Ilho(x)  ho(y)ll, the length of the arc being at least as great as the distance between its endpoints. Therefore Ilx  yll ~ IIho(x)  ho(y)ll. However, a similar argument using hOI rather than ho shows that IIx  yll Ilho(x)  ho(y)ll· Thus Ilx  yll = IIho(x)  ho(y)ll, and ho is an isometry of RP. From this fact, we conclude that h is a similarity transformation of RP. Q.E.D.
:s
We can see the effects of these two types of local shape variation by considering the curvilinear coordinates of Figure 1.7. The coordinate system for the modern human skull was chosen to be a standard Cartesian coordinate system, with intersecting lines meeting at orthogonal and equally spaced parallel lines in each direction. This divides the region into squares. In the curvilinear coordinate systems below this, the images of the squares in the first coordinate system are approximately parallelograms except for those regions where shape variation is occurring too rapidly for the coarseness of
114
3.
Shape Spaces
3.9
the grid. The function M 2 measures those shape changes that stretch the squares of the top grid into the parallelograms of the lower grids. Another type of effect that we can observe is that while the squares in the top grid are all of the same area, the parallelograms of the lower grids have varying areas. This effect is measured by N 2 •
3:8
Notes
Bookstein's approach to shape analysis leads to a manifold of constant negative curvature for the representation of triangle shapes. In contrast to this, Kendall's approach leads to a sphere, which is a manifold of constant positive curvature. This discrepancy need not confuse us nor lead us to consider one geometry superior to the other. In each case, the Riemannian geometry of triangle shape. space is motivated by consideration of the mechanisms that give rise to shape variation. Our generalization of Fred Bookstein's triangle shape geometry to the family of shape spaces UT(n) provides an alternative to the family of spaces ~~+l introduced by David Kendall. Huiling Le has recently computed the anisotropy metric for UT(n). This is the generalization of formula (3.81) from UT(2) to the higherdimensional simplex shape spaces UT(n), where n > 2. The following proposition can be compared with Proposition 3.6.4, which is a special case for infinitesimal distances. Proposition 3.8.1. Let llx and lly be the UTshape representations of x and y respectively. Let A = llyll;l. The square of the geodesic distance from Ux to ll y in UT(n) is given by
t '('~l)n [U IOg(Aj/Ad~IOg(Ak/Ad]2 1)
j=2
J J
115
the average Ofxlx2x3and YIY2Y3 is similar to these triangles. Does this result hold if the triangles are not constrained to lie in the same plane? 2. Show that the 11 correspondence between .~~ and 8 2(1/2) established in formula (3.9) of Section 3.1 is a Riemannian isometry. More specifically, show that formula (1.21) for the distance between shape points on the sphere is equivalent to formula (3.9) for shape distance on ~~. Hint: as both formulas are invariant under similarity transformations, it is sufficient to cohsider two triangles, 1, +1, Zl and 1, +1, Z2, of complex landmarks and to compute the distance between their shapes by the two methods. First find the coordinates of their preshapes and plug into formula (1.21). Then find the coordinates of shape on the sphere from formula (3.7) and plug into formula (3.9). Do you get the same answer?
3. Find all points on the sphere 8 2(1/2) ~ ~~ that correspond to right triangles. What does this region look like? 4. Find all points on the sphere 8 2(1/2) ~ ~~ that correspond to isosceles triangles. What does this region look like? 5. Prove Proposition 3.6.5.
6. Show that left matrix multiplications in UT(2) are not isometries by considering what happens to the matrices and
(
1
o
X+dX) y+dy
(3.107)
under left multiplication by the matrix (3.106)
k=l
where AI,..., An are the eigenvalues of AT A. It can be checked that this reduces to formula (3.81) when n = 2 and to formula (3.72) when y = x + dx. For n > 2 the simplex shape.spaces are not spaces of constant curvature.
3.9
Problems
Problems
1. Consider two similar triangles XlX2X3 and YlY2Y3 in the plane. We define the average of the two triangles to be ZlZ2Z3 where Zl, Z2, and Z3 are the midpoints of XIYl, X2Y2, and X3Y3, respectively. Show that
(3.108) In addition, confirm the results of Proposition 3.6.5 as applied to UT(2) by 'performing the same calculation using right multiplication. 7. Let n ,be the orthogonal n x n matrix defined in Section 3.6.2. Let be the jth column of 3 x . (a) Using the fact that ~j lies in the subspace generated by ~~, ... ,ej, show that nlej has only its first j elements nonzero. (Hint: what does nlej look like?) Conclude that Wx is an upper triangular matrix. (b) Using the fact that < ~j, ~j > is positive, show that the entries down the main diagonal of Wx = n l 3 x are also positive. ~j
8. At the end of Section 3.4, we noted in passing that it is possible to
116
3.
Shape Spaces
arrange a set of points on a Riemannian manifold so that their interpoint geodesic distances do not match the interpoint Euclidean distances of any configuration of points in any dimension. In this problem we shall verify this. Let xl> j == 1, ...,4, be four points spaced at equal intervals around the unit circle S 1. (a) Find the 6 x 6 matrix of interpoint geodesic distances between the points Xj using arc length to measure distance. (b) Show that this 6 x 6 matrix cannot be an interpoint Euclidean distance matrix for any set of four points in any Euclidean space R".
4 Some Stochastic Geometry
4.1 4.1.1
Probability Theory on Manifolds Sample Spaces and SigmaFields
We begin with a review of some basic definitions and ideas from probability theory. The reader wishing a more detailed description of the tools that will be necessary can consult [43]. By a sample space we shall mean a set S whose elements s shall be called outcomes or points. Within S we shall suppose that we have a particular class F of subsets A c S that shall be called events. This class has to be sufficiently rich to allow us to do probability calculations. To do this we shall require that F be a sigmafield of subsets, which we now define. Definition 4.1.1. A class F of subsets of S is said to be a sigmafield on S if the following three properties are satisfied. 1. S E :F. 2. If A E F then its complement N E F. 3. For any sequence AI, A 2 , A 3 , . .. of elements of F the union 00
{s E S : s E A j for some j} is an element of F.
(4.1)
118
4. Some Stochastic Geometry
4.1
Probability Theory on Manifolds
119
Henceforth, we shall assume that the class F of events is a sigmafield. From Definition 4.1.1, it is possible to show that any subset of S constructed as a countable Boolean combination of events in F is itself an event in F.
In the special case where MP = R we also refer to a statistic X on R as a random variable. More generally, a statistic on RP is called a random vector. We can also define a class B of subsets on MP called the Borel subsets.
4.1.2 Probabilities
Definition 4.1.4. The class B ofBorel subsets of MP is defined as the signltafield of subsets of MP generated by the class U of open subsets of MP(
We can now define a probability on a sample space.
\
Definition 4.1.2. By a probability, we mean a function (4.2) such that 0::; P(A) ::; 1 for all A E F
P(S)
satisfying the properties 1
(4.3)
and
ThJs all the open sets of MP are Borel sets including MP itself. In addi\ion, the countable intersection of open subsets of MP is also Borel, although it is, in general, not open. Closed subsets are also Borel, because they are the complements of open sets in MP. The possible types of Borel sets are not exhausted by this list as the possible types of sets generated by countably taking intersections, unions, and complements in any order is very rich indeed. For any Borel set B c MP and for all statistics X S . MP, the set XI(B) = {s E S : X(s) E B}
(4.4)
whenever A j n A k = 0 for all j =1= k. The set S when endowed with a sigmafield F of events and with a probability P is said to be a probability space. If g is any class of subsets of a set S we can also define the sigmafield generated by g. This is the intersection of all sigmafields g' such that g c g'. It is clearly the smallest sigmafield containing the subsets A E g.
4.1.3 Statistics on Manifolds
is an event in S. That is, if B is a Borel set then XI(B) E F. This property indicates the importance of Borel sets in MP. They form a natural class of subsets B for which we can assign a probability that a statistic X lies in B.
4.1.4 Induced Distributions on Manifplds Because the class of Borel sets B is a sigmafield, we can regard the manifold MP as a sample space in its own right, with B as its class of events. A statistic X then induces a probability on B in the same way that the original probability P is defined on :F. We can define the induced probability distribution P XIon the Borel sets of MP to be
Of particular interest are functions from a sample space into a manifold called statistics.
Definition 4.1.3. Let MP be a differential manifold. By a statistic X on MP we shall mean a function
X : S . MP
(4.5)
with the property that XI(U) = {s E S : X(s) E U} is an event (i.e., an element of F) for every open set U G: MP.
(4.8) It can be checked that PXI satisfies the properties of a probability on MP given in Definition 4.1.2 above. In much of probability theory, the fact that statistics are functions on sample spaces tends to be suppressed in the notation. Thus we shall henceforth write X1(B) as (X E B), both being equivalent to the event {s E S : X(s) E B}
(4.6)
(4.7)
(4.9)
So we shall typically write P(X E B) for the probability in equation (4.8). Other notations are similar. For example, if MP = R , we write (X::; t) for the set (X E (00, t]), etc. Another abbreviation is to use a comma to
120
4.
4.2
Some Stochastic Geometry
stand for the logical operation "and" and the corresponding set operation of intersection. Thus we shall write
(4.10) to stand for (Xl E B I ) n (X2 E B2)' We can make new statistics out of old. One way to do this is through the use of Cartesian products. For example, if Xl is a statistic on MP and X 2 is a statistic on Nq then X = ( Xl, X 2) is a statistic on MP x N". Correspondingly, any statistic X on MP x N? defines Xl and X 2 uniquely on MP and N? respectively. Another way to build new statistics out of old ones is through composition of functions. For example, if X is a statistic on MP and h : MP > N? is continuous, then h(X) : S
>
N?
(4.11)
is a statistic on N". This follows from the fact that the continuous preimage hl(B) of a Borel set Be N" is a Borel set of MP.
The Geometric Measure
121
4.1.6 Stochastic Independence We now briefly review some basic definitions and properties related to stochastic independence. Let Xl, X 2 , 00" X n be statistics taking values in a differential manifold MP. These statistics are said to be mutually stochastically independent, or simply independent, if for all Borel sets B I, B 2 , •.. , B n c MP we have n
II P(X
j
E Bj )
(4.14)
j=l
If the induced distributions P x;' on MP are equal in the sense that
P(X I E B)
= P(X2
E B) =
00.
= P(Xn
E B)
(4.15)
for all Borel B C MP then we shall say that Xl,. 00' X n are identically distributed. The condition that random variables are both independent and identically distributed is often abbreviated by saying that they are IID.
4.1. 7 Mathematical Expectation
4.1.5
Random Vectors and Distribution Functions
Suppose X is a random vector taking values in Euclidean space RP. If we write X in terms of its coordinates Xl,oo.,X p then Xl,oo.,X p are random variables. We define the distribution function of X to be a real valued function Fg : RP > R such that (4.12) It is important and nontrivial to show that the induced distribution P XI on the Borel sets of RP is determined by its joint distribution function Fx. We say that It random variable X is absolutely continuous, or simply continuous, if there exists a nonnegative function f : R > R such that P(X E U)
=
1
f(x) dx
(4.13)
xEU
for all open sets U C R. Equation (4.13) holds true if U is replaced by any Borel subset of R. A random vector X taking values in RP is said to be absolutely continuous if the higherdimensional analog of (4.13) holds for some nonnegative function f: RP > R and all open U C RP. Many probability distributions can be constructed on RP that are not continuous, although many of the models in this book will be of the continuous type. Another important class of probability distributions are the discrete probability distributions, which assign unit probability to some countable set.
If X is a random variable with distribution function F, then we can define the mathematical expectation, also known as the mean or expected value of X, to be + 00
£(X) =
1
00
x dF(x)
(4.16)
for those random variables for which the integral is finite. The expected value of a random vector X = (Xl,oo.,Xn ) we shall define as the vector of expected values (4.17) £(X) = [£(Xd, 00', £(Xn ) ] In a similar way, the expected value of a matrix can be defined as the matrix of expected values of its elements. We cannot do justice in this brief survey to the full range of definitions and results on expectation, independence, conditional probability, and marginalization of distributions. The reader is referred to standard sources for the .results needed throughout the remainder of this book.
4.2
The Geometric Measure
We now seek to generalize the concept of a pdimensional content, or volume, in RP to a Riemannian manifold. We have already seen that the metric tensor is instrumental in defining the lengths of paths in a Riemannian manifold. One would naturally expect the metric tensor to be essential
122
to the definition of content as well. We begin by considering the relationship between the metric tensor and pdimensional content in RP. Let (4.18) for j = 1,2, ... ,p be a set of p linearly independent vectors in RP. The basis vectors Xl, X2, ..., x p define the edges of a parallelepiped in RP given by
{
aj Xj : O:S aj:S 1 for all
t 3=1
j}
The volume function Vp , when extended to the Borel sets of the manifold, is called the geometric measure. Now suppose S is a probability space endowed with a probability P, and suppose X : S + MP is a statistic on the Riemannian manifold. The statistic X will be said to be absolutely continuous, or simply continuous, provided that there exists a nonnegative function f: MP + R such that P(X E U) =
r
f(x) dVp(x)
for all open sets U on the manifold. If this is the case, we shall call f the density function of X. An important special case occurs when Vp(MP) < 00 and (4.27)
X12
X22 (4.20)
(
Xpl Xp2 Now, Problem 4 asks the reader to show that
where
P
gjk = < Xj,Xk > = EXjlXkl
In this case, we say that X is uniformly distributed on the manifold, or that the induced distribution on MP is uniform. Note that the density can never be constant on a manifold for which Vp(MP) = 00.
4.2.1 (4.21)
(4.22)
Example: Surface Area on Spheres
To illustrate the idea of volume and content, consider the 2sphere S2(r) of radius r from the example in Section 2.2.14. Let Ih be the longitude and (}2 the colatitude of that example as defined in formula (2.63). Then applying formula (4.24) above, we see that
1=1
is the metric tensor for RP endowed with a coordinate system based on Xl, X2, ..., x p rather than the usual orthonormal set of vectors. Formula (4.21) tells us that pdimensional volume is characterized by the metric tensor. To calculatepdimensional volume on a general Riemannian manifold we use an approach that is similar to the calculation of arc length using the metric tensor. The metric tensor on a Riemannian manifold MP permits us to define the volume of a parallelepiped in Tx(MP), the tangent space at X E MP. From this we define the volume dVp(x) of a small region whose coordinates are
{y
Xj:S Yj :S Xj + dXj for all j}
= (Yl' Y2, ..., Yp) E MP
(4.23)
to be
dVp(x)
=
(4.26)
}XEU
(4.19)
A standard result in linear algebra tells us that the volume, or content, of this parallelepiped is the absolute value of the determinant of the matrix X ll .X21
123
4.2 The Geometric Measure
4. Some Stochastic Geometry
(4.28) This formula is quite commonly derived from heuristics in multivariate calculus courses. Thus the surface area of the sphere is (4.29) as is well known.
4.2.2
Example: Volume in Hyperbolic Half Spaces
Consider the hyperbolic half spaces of Section 2.2.17. For these spaces (4.30)
jdet(gjk dXjdxk)
(4.24)
The volume of an open set U in the manifold is then found to be
Vp(U) =
1
xEU
dVp(x)
(4.25)
Close to the horizon at infinity where x p = 0, when measured in Euclidean coordinates the volume element dVpgoes to infinity. Unlike their positive curvature counterparts, the spheres SP, the hyperbolic half spaces HSP have infinite volume.
124
4. Some Stochastic Geometry
4.3
Transformations of Statistics
4.3.1
4.4 Invariance and Isometries
Jacobians of Diffeomorphisms
Consider two manifolds MP and NP of the same dimension, and let h : MP  t NP be a differentiable function. Suppose MP and NP have coordinate systems x = (Xi""'X p) and Y = (Yi, ... ,Yp) respectively. Then we can express h in terms of these coordinates as a function
125
using formula (4.24). In measuretheoretic language, we can also call the ratio of differentials a RadonNikodym derivative. An extension of (4.34) allows us to calculate the distribution of Y = heX) for transformations h : MP  t N? where q < p. Suppose we can find a manifold NPq and a transformation (4.36) such that
(4.31) Then the Jacobian matrix of h can be defined in terms of these coordinates as the matrix of partial derivatives A = (8Yjj8xk) as in formula (2.9). Similarly, we can define the Jacobian of h to be (:Jh)x
=
det (A)
(4.32)
Note that the Jacobian is dependent on the particular coordinate system used on each manifold. That is, the Jacobian is extrinsic to the manifolds. However, it appears in calculations to compensate for changes of coordinates, and therefore can be used to build intrinsic quantities that are independent of the coordinate system. If h is a diffeomorphism between manifolds then (4.33) Moreover, both quantities will be bounded away from zero and infinity. If h is a local diffeomorphism, the identity (4.33) remains true because of the Jacobian's local nature. But in this case, the inverse transformation has to be interpreted locally.
4.3.2
Change of Variables Formulas
Now suppose X is a statistic on MP with density function f and suppose that h : MP  t NP is a diffeomorphism. We define the statistic Y = heX), and consider the problem of calculating the density function of Y on NP. Let gM be the metric tensor on MP, and let gN be the metric tensor on NP. Then it can be shown that the density function of Y on NP can be written in terms of the Jacobian :Jh i and the metric tensors as (4.34) where the ratio of differentials of the geometric measure is a coordinatefree notation for the expression det gM[hi(y)]  detgN~
(4.35)
(h, h') : MP
t
N?
X
NPq
is a diffeomorphism. Let hi = (h, h') and Y' function of Y1 = hi (X) = (Y, Y')
(4.37) h'(X). The density
(4.38)
can be calculated from formula (4.34) above. We then find the marginal density of Y = heX) by integrating this density over its second variable, leading to the formula (4.39)
4.4
Invariance and Isometries
In order to prove that a particular induced distribution on MP is uniform, it is possible to check directly that its density is constant. Often, however, there is another way based upon the concept of invariance. Suppose that for any two points x, y E MP there is a geodesic from x to y. Then there is a well defined concept of geodesic distance between points. For such manifolds, we can use Iso(MP), the group of isometries on MP, to investigate whether a statistic has a uniform distribution on MP. Now the volume measure on MP is invariant under isometries in the sense that
Vp [h(B)] = Vp(B)
(4.40)
for all Borel sets B and for all h E Iso(MP). Similarly, if X has a uniform distribution on MP then the probability distribution is invariant under the group in the sense that
P[X E h(B)] = P(X E B)
(4.41)
This invariance property of (4.41) is illustrated in Figure 4.1. We shall be concerned with the converse of this result. Namely, if a continuous statistic X has this invariance property with respect to Iso(MP) does it follow that X is uniformly distributed on MP? The answer, in general, is no.
126
4:4, Invarianceand Isometries
4. Some Stochastic Geometry
127
h is an isometry, it follows that the Jacobian matrices of hand h 1 will be orthogonal at the points x and y respectively. Therefore,
.~
~y
(4.42) Furthermore, because X has a distribution that is invariant for all h E Iso(MP), the statistics X and heX) have the same distribution and the same density functions. Thereforeformula (4.34) gives us
(4.43) for all y. An Immediate consequence of this is that f must be a constant function when Iso(MP) is transitive. Thus X must have a uniform distribution. Q.E.D. FIGURE 4.1. Invariance of the uniform distribution with respect to isometries.o] the manifold. A Borel set B is shifted by an isometry h. of the manifold to a set h(B). If X is a uniformly distributed statistic on the manifold, then X will lie with equal probability in Band h(B).
However, for an important special case, the converse is true. The following definition provides the necessary class of isometries for which a converse can be obtained. Definition 4.4.1. Let H be any group of transformations on MP. Then H is said to be transitive if for every x and y in MP, there is an hE H such that hex) = y. Using the concept of transitivity of the group action on MP, we can obtain our converse. Proposition 4.4.2. Let X be a continuous statistic on the manifold MP, with density function f(x). Suppose that H is any transitive subgroup of Iso(MP), and that the distribution of X is invariant under H in the sense of equation (4.41) above. Then f is a constant density on MP. That is, X is uniformly distributed on MP. Proof. This result follows from a special case of formula (4.34), in which MP = NP and gM =gN' To prove that X is uniform, we must show that f(x) = fey) for all x,y E MP. As H is a transitive group, it follows that there exists an hE H such that hex) = y. We can set up coordinate systems around x and y so that aM (x) and gN(y) are the identity matrices. (This can be achieved at specific' points in a manifold, but will only hold over an open set when the manifold is flaton that set.) Because
4.4.1
Example: Isometries of Spheres
In 4.2.1 we wrote the differential dV2 of surface area in terms of the coordinate system. We now characterize surface area through invariance. Both O(p) and U(p) are isometry groups on RP and CP respectively. Now O(p) maps the unit sphere Spl to itself, and is therefore an isometry group for Spl. It is a simple exercise to show that this group acts transitively on Spl. See Problem 6 at the end of the chapter. The uniform distribution is the unique distribution on S2 that is invariant under the action of 0(3). More generally, the uniform distribution on Spl is the unique invariant distribution under the action of O(p). Now, the unitary group U(q) is identifiable as a subgroup of O(2q). Thus U(q) is a group of isometries of S2ql. With a bit of work, we can also show that U(q) acts transitively on S2ql. See Problem 7. Thus the uniform distribution on S2ql Is also characterized as the unique distribution that is invariant under U(q).
4.4.2
Example: Isometries of Real Projective Spaces
Let
(4.44) denote the covering mapping taking each x E SP to the pair of antipodal points {x, x} E RpP. Suppose X is a statistic that is uniformly distributed on SP. As A is continuous, it follows that A(X) is a statistic on RpP. As will be seen, A(X) is uniformly distributed on RpP. The fact that this is true can be shown by calculating the density function of A(X) directly. If X has density function f on the sphere SP, then it can be shown that the density function of A(X) at the point A(x) is f(x) + f( x). The constancy of f on SP implies the constancy of f(x) + f( x), and thereby the uniformity of the distribution of A(X).
128
4. Some Stochastic Geometry
4.4 Invariance and Isometries
However, an alternative proof using invariance is useful. To prove the result, we first note that an orthogonal transformation of the sphere preserves the property that points are antipodal. That is, h(x)
=
h(x)
(4.45)
for all transformations hE O(p+ 1). This means that the group O(p+ 1) acts on RpP as well, the element h E O(p+ 1) taking any pair of antipodal points {x, x} to another pair of antipodal points {h(x), h(x)}. At the risk of some confusion, we write (4.46)
129
the first equality following from the invariance of the uniform distribution on the sphere SP. This demonstrates the invariance. So A(X) is uniformly distributed on RpP. The reader should note that the group O(p + 1) is a little bigger than the group of isometries it induces on RpP. The transformation x  t x is an element of O(p + 1) that maps {x, x} back onto itself. Thus it induces the identity transformation on RpP. This transformation, together with the identity transformation, forms a subgroup of O(p + 1) that is the center of O(p+ 1). It is isomorphic to the group 0(1). Thus in grouptheoretic terms we can write the group of isometries induced on RpP as the factor group O(p + 1)/0(1).
as well as (4.47)
letting the context decide the transformation under consideration. With this understanding, we can show that A[h(x)] = h[A(x)]
(4.48)
for every hE O(p + 1) and every x ESP. More compactly, we can write A 0 h = h 0 A. Equivalently, the functions A and h can be said to commute. The diagram of this looks as follows:
(4.49)
Now O(p + 1) can be checked to be a group of isometries of RpP, because the mapping {x, x}
t
{h(x), h(x)}
E
h(B)]
P {X
E
Al[h(B)]}
{X
E
h[A1(B)]}
(4.51) P
The second equality follows from the fact that h and A However, P [X E A1(B)]
commute.
(4.52) P[A(X) E B]
Example: Isometrics of Complex Projective Spaces
Our next example is an extension of the previous case to include the class of complex projective spaces. Suppose that Z is a statistic that is uniformly distributed on the sphere S 2q +l , this time understood as the unit sphere about the origin in C q + 1 . Let (4.53) be the mapping taking each point z of the sphere S2q+l into its orbit O(z), as in Section 2.2.16. We claim that O(Z) is uniformly distributed on cPq. The proof of this result parallels the case for RpP above, with the exception that thegroup U (q + 1) must be used to provide the invariance rather than O(2q + 2). The proof goes through in a similar way to that of Section 4.4.2, above. In this case, our commutative diagram becomes
(4.50)
preserves geodesic distance in RpP. That O(p + 1) acts transitively on RpP follows easily from the fact that it acts transitively on SP. See Problem 6 at the end of the chapter. In addition, for any Borel set Be RpP,
P [A(X)
4.4.3
(4.54)
The transformations on cPq induced by U(q + 1) are isometries, the group as a whole acting transitively on cPq. As above, this follows easily from the fact that U(q + 1) acts transitively on S2q+l. See Problem 7. The reader should note that as in the previous example, the group U(q+ 1) is bigger than the group of isometries it induces on cPq. A subgroup of U(q + 1) determined as its center maps orbits O(z) back to themselves. Such transformations induce the identity transformation on cPq, and together form a group that is isomorphic to U(l). In a manner similar to the previous example, in grouptheoretic terms we can write the group of isometries induced on cPq as the factor group U(q + l)/U(l).
130
4. Some Stochastic Geometry
4.5 Normal Statistics on Manifolds
4.5 Normal Statistics on Manifolds
4. 5.1
131
first j 1 entriesequal 1/)j(j 1), the jthentry being )(j 1)/j, and whose remaining nj entries are zero. Thus, for example, the Helmert matrix of order 4 is
Multivariate Normal Distributions
In this section, we give a brief summary of some definitions and results from multivariate normal theory for Euclidean spaces and spheres that we shall need for shape modeling. In keeping with the spirit of Section 4.4, we shall consider these models from the perspective of invariance. Definition 4.5.1. Let X = (X I,X2 , ... ,Xn ) T be a column vector of random variables. Then the random vector X is said to have a multivariate normal distribution if it has a density function of the form
1/..;4
1/..;4
1/..;4
1/..;2
1/..;2
0
o
1//6
1//6
2//6
0
(4.56)
1/V'12 1/V'12
1/V'12 3/V'12
(4.55) where J.L = (J.LlJ J.L2, ... , J.Ln)T is a column vector, called the mean vector, and r is an (n x n)dimensional positive definite symmetric matrix,called the covariance matrix. In the special case where n = 1, we simply say that X has a normal distribution with mean parameter J.L and variance parameter r > 0, understanding X, J.L, and r to be scalars.
In fact, J.L = £(X) and r = £(XX T ). The entries X j of a multivariate normal vector X can be shown to be normally distributed random variables. Suppose that X > AX +a is an affine transformation of full rank, where a is an n x 1 column vector. Then it can be shown that Y = AX + a also has a multivariate normal distribution, with mean vector AJ.L + a and covariance matrix Ar AT. Of particular interest to us here will be the special case where r = cI. In this case, we say that X has a spherical normal distribution. The random vector X can be shown to have a spherical normal distribution if and only if the random variables Xl, X2, ..., X n are independent normal random variables with common variance parameter. The spherical normal density function is preserved under similarity transformations of R". Suppose X is spherical normal and Y = bAX + a, where A is an n x n orthogonal matrix, a is an n x 1 column vector, and b is a positive scalar. Then Y is also spherical normal.
4.5.2 Helmert Transformations The following class of orthogonal transformations and their matrix representations will be of interest for shape theory. By a Helsnert matrix of order n we shall understand an n x n matrix whose first row is a row vector of entries equal to 1/;n. For j = 2, ..., n, the jth row is a row vector whose
The Helmert matrices can be shown to be orthogonal. Now suppose that X I, ... , X n are independent identically distributed random variables. Then X = (Xl, ..., Xn)T has a spherical normal distribution. Suppose also that A is a Helmert matrix of order n. As A is orthogonal, it follows that Y = AX is also spherical normal. This implies that Y I , Y 2 , ... , Yn are independent normal random variables with common variance. With the exception of YI , which generally has nozero mean, the other random variables Y2, ..., Yn have zero mean, and are therefore identically distributed. The random vector (1'2, ..., yn)T can be placed in onetoone correspondence with the vector of residuals (Xl  X, ..., X n  X) used to eliminate location information from landmarks in Chapter 1. The former vector can be regarded as an orthogonalized form of the latter, which has a linear constraint on its components.
4·5.3 Projected Normal Statistics on Spheres As we saw in Chapter 1, projection onto a sphere arises in shape analysis from the removalof scale variables in the reduction to the preshape of the data. Definition 4.5.2. Suppose that X E R" has a spherical normal distribution with mean vector J.L. With probability one, X will be nonzero. Let IIXII be the norm of X. The scaled vector B(X)
X
IIXII
(4.57)
is a point on the unit sphere snl centered about the origin. Thus B(X) is the projection of a normally distributed vector onto snl and is said to have a projected normal distribution.
132
4.
4.5 Normal Statistics on Manifolds
Some Stochastic Geometry
To compute the density function for the projected normal distribution on sn1, we transform variables, writing any nonzero point x E R" in polar form as (r,B), where r = Ilxll and B = x/llxll. The pair (r,B) naturally lies in the product manifold R+ X sn1. Under the identification of x with (r, B), the volume element on R" decomposes as
(4.58)
Without loss of generality, we consider the standardized case where the covariance matrix of X is the identity. If this is not the case, then X can be rescaled before projection onto sn1. Let Bj be the angle between B and the xjaxis. So Xj = r cos(Bj ) . Then the density function with respect to the element dVnl (B) dr is (4.59)
Thus the density function of B with respect to the volume element dVn  1 ((;I) has integral representation (4.60)
For simplicity, let us assume a coordinate system where the components of the mean vector of X have the form /kl = v and /k2 = .., = /kn = O. Then the density function reduces to
Following the notation of [70], we define
Tk(u) =
LX)
r
k exp ( ~;)
«» dr
The special functions can be computed recursively using the formulas To(u) = ~(u)!¢(u)
(4.64)
+
(4.65)
=
1
'U~(u)!¢(u)
(4.66)
where if> is the density function for a standard normal random variable and ~ is the distribution function of a standard normal random variable. For a more general sequence {L = (/kl,/k2, ... ,/kn)T we replace B1 in formula (4.60) by the angle made between the vector B and the vector ts, Problem 9 asks the reader to investigate the density function when v = 0 and u > 00. It can be seen that when v = 0, the distribution on snl reduces to the uniform distribution, the density having a value that is the reciprocal of the volume of the sphere. To see the relevance of this distribution to shape analysis, and the distribution of preshape statistics in particular, consider a set of independent random vectors (4.67) each of which has a spherical normal distribution. Suppose also that the random vectors have common pxp covariance matrix el, say, but possibly different mean vectors (4.68)
We construct a p x n matrix X whose jth column is X j . The matrix X can also be regarded as a vector in Rpn whose components are the entries X j k. Let A be the (n  1) x n matrix made by deleting the first row of the Helmert matrix of order n. Then the p x (n  1) matrix (4.69)
has a spherical normal distribution in RP(nl). The columns of the matrix Yare independent and identically distributed normal random variables containing the information from X with location information removed. We find the preshape of X by rescaling Y so that the sum of squares of its p(n  1) components equals one. This is equivalent to projecting Y as a vector in RP(nl) onto the unit sphere snpp1 around the origin in RP(n1). Let
Y T
(4.63)
Ll(U)
and
(4.62)
for k = 0, 1,2, .... Then the density function can be written as
133
=
IIYII
E
snppl
(4.70)
be the preshape of X. Then T is seen to have a projected normal distribution on snppl. The reader should note that although for the case p = 2 the coordinate representation of preshape information given here differs from the representation of Chapter 1, the two representations are isometric. In the earlier chapters, the sphere s;n3 was embedded as the unit sphere in a (2n  2)dimensional subspace of R 2n . However, here we use the sphere s2n3, which is the unit sphere in R 2 n  2 . The reader can demonstrate
134
4. .Some Stochastic.Geometry
4.6 Binomial and Poisson Processes
the isometry by checking that the formulas for geodesic distances between preshapes are identical for the two representations. To go from preshape statistics to shape statistics involves one additional integration. In view of the complexity of the projected normal density function, it might be questioned as to whether the associated shape density can be written in manageable form. It was the conclusion of Mardia and Dryden [116] that for p = 2, this distribution is not only simple in form but can be written in terms of the geodesic distance between shapes on the shape manifolds ~~ ~ Cpn2. We will consider this in detail in the next chapter.
Let B be an open subset of a Riemannian manifold MP such that Vp(B) < 00. Now, suppose that B o is an open subset of B. Let Xl, X 2 , ••• , X n be n independent statistics in MP that are uniformly distributed over the open set B. We define the nonnegative integervalued random variable N to be n
N The random variable N with the property that
Binomial and Poisson Processes
Uniform Distributions on Open Sets
We have noted that if MP has infinite volume, then no uniform distribution exists for it. However, it is possible to impose a uniform distribution on regions of the manifold and to associate with these uniform distributions a limited form of invariance. Let B be an open set of MP for which Vp(B) < 00. A continuous statistic X E MP is said to be uniformly distributed on B if its density function has the form
f(x) =
VPCB. ) {
o
xEB
(4.71) xtfB
A particular case of interest to us will be thatfor which MP = RP and B is a convex subset of RP. We shall explore this in greater detail in the next chapter, where the formulas of integral geometry will be directly applicable to the shapes of points uniformly and independently generated in convex sets.
4.6.2
L
lcxjEBo)
(4.74)
is well known to have a binomial distribution,
(n) [Vp(Bo)] k [1 Vp(Bo)]nk Vp(B).· Vp(B) k
We now turn to some examples of point processes that will be important for our development of the statistics of shape.
4.6.1
=
j=l
peN = k) =
4.6
135
(4.75)
For this reason, an indepedent and uniform scattering of a fixed number of statistics over an open set B is often called a binomial process. For a binomial process, the random set of points in MP so generated is the principal object of interest, the ordering of the points being of secondary interest.
4.6.3
Example: Binomial Processes of Lines
Line processes provide a mathematical model for the random scattering of straight lines. Examples include the cracking of surfaces and the tracks left by small particles scattered randomly and homogeneously through a region with random orientations to their velocities. Suppose X is a line in the plane that passes through a bounded convex set A C R 2 . Let us suppose that X is directed. We can think of this as providing a rule as to which is the left side of the line. The rule is arbitrary but must be consistently imposed all the way along the line. The set of all directed lines in the plane can be placed in a natural 11 correspondence with the points of the cylinder R x 8 1 , as shown in Figure 4.2. The group Euc(2) of Euc!idean motions of R 2 maps lines to lines. Thus Euc(2) can be said to act upon the manifold R x s. A Euclidean motion takes a line with coordinates (r,O) to its image with coordinates (r', 0'). Among these transformations are the translations on R x s' mapping (r, B) > (r+s, 0)
Binomial Processes
(4.76)
and
For any event A the indicator random variable
(r, 0) (4.72)
for the event A is defined by (4.73)
>
(r, 0 + 0 when the origin is on the side of the line as shown and r < 0 when the origin is on the other side.
137
is invariant under these translations. Moreover, the group of translations generated by (4.76) and (4.77) acts transitively on R x s. Therefore, we can show in a manner similar to Proposition 4.4.2 that the measure in (4.78) is the unique invariant measure up to an arbitrary scalar multiple. See, for example, Santalo [147, pp. 2730] for the details of this. Using dV2(r,l1) we can construct a binomial process of directed lines passing through A. Let B be the set of all (r,l1) such that the line with coordinates (r,l1) passes through A. As V2(B) < 00, we can construct a collection of independent random lines Xl,. '" X n that are uniformly distributed in B. Having constructed such a binomial process of lines, we can make the lines undirected by erasing the arrows. We can also consider the shape of any configuration of lines generated by a binomial process. Like Euc(2), the group Sim(2) maps lines to lines, and therefore can be regarded as acting upon the manifold R X 8 1 . The shape of any configuration of lines can be defined as the total information in the set Xl, "., X n that is invariant under the action of this group. The situation is analogous to our definition of the shape of a set of landmarks in Chapter 1. However, instead of the direct action of 8im(2) on R 2 , the action of Sim(2) on the space of lines is induced from the action on the Euclidean space in which the lines reside. In this and other contexts, the study of the shapes of configurations of geometric objects becomes the study of invariants of configurations of points in manifolds. See Carne [38]. The assignment of directions to lines is a mathematical convenience that allows us to put a simple coordinate system on the space of lines. Undirected lines are more physically natural for the purposes of modeling many physical processes involving line data. We obtain an undirected line by throwing away the direction, so to speak. More formally, we can define an undirected line as a pair of directed lines having coordinates of the form {(r, B), (r, O+7l')}. The manifold of undirected lines in the plane is thereby seen to be the space of such pairs of points in the manifold R x 8 1 . Wilfrid Kendall has noted, in a private communication, that the space of undirected lines is homeomorphic to the Moebius strip defined in Problem 3 of Chapter 2. Problem 11 at the end of this chapter sketches the steps necessary to prove this fact.
4.6.4 Poisson Processes Let us return to the general binomial process in some open set B c MP. Suppose that Vp(MP) = 00. A limiting case is obtained when B expands to encompass all of MP. As noted earlier, there is no uniform distribution on all MP in this case. However, there is a nondegenerate limiting form for the binomial process. Consider a nested sequence of open sets (4.79)
138
4. Some Stochastic Geometry
4.7 Poisson Processes in Euclidean'Spaces
so that MP = U~lB.j' Suppose that as B j of positive integers
/'
RP we select a sequence
139
Definition 4.6.3. A VOlumepreserving point process on MP is said to be a homogeneous Poisson process if it satisfies the following postulates.
(4.80) such that
Postulate 1. If 0 < Vp(B) < 00 then 0 < P[N(B) as Vp(B) + 0 we have P[N(B) = 0] > 1.
n·
Vp(~j)
>
p> 0
(4.81)
1. Moreover,
Postulate 2. If the sets B l , B 2 , ••. , B m , m 2: 2, are disjoint subsets of RP then N(Bd, N(B2 ) , .•. , N(Bm ) are independent.
as j + 00. For each j we construct a binomial process of nj points in B j , and for each we let N j be the number of such points falling into B o. Then PEN. = k) + [pVp(BO)]k exp[pVp(Bo)] (4.82) J
= 0]
1] > 0 P[N(B) = 1]
for k = 0,1,2, ..., which is the well known formula for the Poisson, distribution. The limiting form of the binomial process is called the Poisson process oj intensity p. See Problem 10 for the derivation of the Poisson formula. We can think of the Poisson process intuitively as a uniform scattering of infinitely many points throughout the entire space, so that on average, p points fall into a region of unit volume. We shall need to refer to given points of MP that are distinct from the random set of points of the Poisson process. Someterminology helps to keep these distinct. We shall henceforth refer to the random points of a point process as particles and shall reserve the term points for given elements of MP that have a fixed location. However, following traditional terminology we shall continue to refer to a random scattering of particles as a point process. The following definition helps to formalize the construction ofthe Poisson process by characterizing it in terms of its properties.
For a Poisson process (i.e., a point process satisfying the above), it can be shown that there exists a unique intensity parameter p > 0 such that for every open set B c MP the random variable N(B) has a Poisson distribution with parameter pVp(B) as given in formula (4.82) above.
4.7
Poisson Processes in Euclidean Spaces
In this section, we will summarize some of the properties of Poisson processes in pdimensional Euclidean space. The strong invariance properties of the Poisson process make it a particularly useful model for generating random shapes. The invariance of shape statistics under Euclidean motions is compatible with the motioninvariance of the Poisson process, making probability calculations easier. Although the assumptions of the Poisson process will not typically be realized in their exact form in applications, the model has been found to be useful for simulating a variety of phenomena involving random scatterings of particles.
Definition 4.6.1. We define a point process on MP to be a random countable set of particles C c MP. A point process that has finite intersection with any bounded subset of MP is said to be locally finite. Henceforth, we shall assume that all point processes under consideration are locally finite. For any bounded Borel subset B, let N(B) be the cardinality of the set B n C. Among the class oflocally finite point processes are those satisfying certain uniformity conditions as given in the next definition.
4.7.1
Definition 4.6.2. A point process is said to be volumepreserving if the N(B l ) and N(B 2 ) are identically distributed whenever Vp(B l ) = Vp(B 2 ) .
Among the class of volumepreserving point processes are the Poisson point processes that we described above as the limit of binomial point processes. We have the following definition:
(4.83)
:.j"
.J
Nearest Neighbors in a Poisson Process
A number of geometric properties of the particles of a Poisson process (PP) hold with probability one. For example, a nearest neighbor of a point x E RP is a particle X of the PP that has minimum distance from x among all such particles of the PP. It can be seen that with probability one every particle of the PP has a unique nearest neighbor. This can be generalized to the second nearest neighbor, and so on. In general, the kth nearest neighbor of a point x is a particle X such that there are exactly k  1 particles of the PP strictly closer to x than X. Again, with
140
4.
4.7 Poisson Processes in Euclidean Spaces
Some Stochastic Geometry
141
sphericity property of the PP in RP states that with probability one such a sphere passing through p + 1 particles will meet no other particles of the PP. Particles may be found in the pdimensional ball bounded by the sphere, but not on the spherical boundary itself.
, , , : , , ' , , ,, , ,, ,, , ,
, , , ,
, ,, '
., ,, ' , ,, ,
x
,
, r
X3
FIGURE 4.3. The nearest neighbors in a point process. Given any fixed point x, with probability one no two particles of a Poisson process will be at the same distance from x. The closest particle to x, labeled Xl in the diagram, is called the nearest neighbor of x, and is on a sphere centered about x that has no particles in its interior. In general, the kth nearest neighbor of x, labeled Xk, is on a sphere centered about x with k  1 particles in its interior. This fact implies that the distance to the kth nearest neighbor of x will be greater than s > 0 if and only if there are k  1 or fewer particles in the interior of the sphere of radius s centered at x. This property can be used to calculate the distribution function of the distance to the kth nearest neighbor of z , . probability one the kth nearest neighbor is unique. Note that the point x can itself be a particle of the PP if desired, with the understanding that it is not its own nearest neighbor. See Figure 4.3.
4.7.2
The Nonsphericity Property of the PP
A set of p + 1 particles of a PP in RP are said to be in general position if the convex hull of the particles has an nonempty interior (or equivalently, contains an open subset of RP). Thus three particles are in general position in R 2 provided they are not collinear. Four particles are in general position in R 3 provided no three are collinear and the four particles are not coplanar, etc. It can be shown that with probability one for a PP in RP, all sets of p + 1 particles of the PP are simultaneously in general position. A property related to this is the nonsphericity property. Through any set of p + 1 particles, which with probability one are in general position, a unique (p  I)dimensional sphere may be drawn, for p 2: 2. The non
4.7.3
The Delaunay Tessellation
There are many mechanisms for selecting finitely many points from a Poisson process for the purpose of generating shape distributions. One of the most important is the Delaunay tessellation, which decomposes RP into pdimensional simplexes (i.e., triangles, tetrahedra, etc.) that are nonoverlapping in the sense that any two simplexes can share at most a common (p  1)dimensional face. The vertices of these simplexes are the points of the Poisson process itself. To construct the Delaunay tessellation, we take advantage of the nonsphericity property of the Poisson process. We have the following definition:
Definition 4.7.1. Let Xl, ... , X p +1 be a set of p+ 1 particles from some PP in RP. Let (4.84) be the pdimensional simplex whose vertices are these p+ 1 particles. We say that ~ is a Delaunay simplex of the PP provided that the (p  1)dimensional sphere passing through X 1,X2 , ""XP+l encloses no particle of the process within its interior.
For the planar case the simplexes are triangles, and so we speak of the Delaunay triangles. Figure 4.4 shows the Delaunay triangles associated with a particular arrangement of particles in R 2 . We now have the following:
Definition 4.7.2. A collection {~j} of countably many pdimensional simplexes in RP is said to be a tessellation if Uj ~j = RP and if in addition, the interiors of the sets ~j and ~k have empty intersection whenever j f= k. We state the following proposition without proof. The reader who is interested in the details of the proof should consult [128].
Proposition 4.7.3. With probability one the Delaunay simplexes of a PP in RP form a tessellation. In fact, a stronger statement can be made about the tessellation. Any two Delaunay simplexes with nonempty intersection will share a face in
4.7 Poisson Processes in Euclidean Spaces
4. Some Stochastic Geometry
142
143
common, the face being itself a simplex, of dimension p  1. The Delaunay tessellation of a PP provides a stochastic mechanism for generating simplex shapes. We shall consider this shape distribution in the next chapter. At this stage, we consider the distribution of the geometric characteristics of a typical random Delaunay simplex. The precise nature of this distribution depends of course on what is meant by a typical simplex. Along with the randomness of the PP that generates the tessellation, the idea of a random Delaunay simplex presupposes a stochastic mechanism for selecting a cell from the tessellation.
I~=::::::::::::~
,r , , ,, , \
4.7.4 PreSizeandShape Distribution of Delaunay Simplexes ,,
,, ,
,
,,
, \
,
, \
\
,
,
,
,
,
\
, ,
\
I
Consider a binomial process of n 2:: p+l independent particles Xl, ..., X n that are uniformly distributed in an open convex subset A of RP. Given that the simplex ~ with vertices Xl, ..., Xp+l, say, forms a Delaunay simplex, what is the distribution of the joint geometric characteristics of the simplex? For ti to be a Delaunay simplex, we require that none of the points X p +2 , ... , Xn fall inside the circumsphere through Xl, ..., X p + 1 ' Let
\
I
I
(4.85)
I I
,
I I
,, ,, ,
I
I I
I I
, ,
I
\ \
,
,,
r \ ,, , ,
"
, I
.........
FIGURE 4.4. Delaunay triangles in the plane. With probability one the particles of a PP satisfy the nonsphericity property. Therefore the circumcircle through any three particles will meet no other particle of the PP. However, many such circumcircles will have particles in their interiors. Those triangles whose circumcirclesdo not have any particles within their interiors are called Delaunay triangles. If the PP is generated to fill the plane, then the Delaunay triangles form a tessellation of the plane.
be the radius of the circumsphere through X}, ..., Xp+l' For large n, the case of particular importance here, the volume enclosed by the circumsphere will be small compared to the volume of A, because for other cases a particle will lie with high probability in the ball enclosed by the circumsphere. It follows from this that in the conditional case that ti is a Delaunay simplex, with high probability as n . 00 the circumsphere will lie entirely within A. For the probability calculations that follow we shall restrict to those cases where the circumsphere lies entirely within the interior of A. Given ti, the expected number N of particles among X p + l , ... , X n that fall into the interior of the circumsphere through ~ is c(N)
[n  (p + 1)] P(X E
~O)
(4.86)
[n  (p + 1)] r P K:p/Vp(A) where X is any of the particles X p +1 , ... , X n and K: p is the volume of the unit ball in Rp. Given ti, the number of particles Xl'+2, ..., X n that fall inside this circumsphere has a binomial distribution. So the Poisson approximation for the probability that no particles among X p +1 , ... , X n fall inside the circumsphere is  00. 10. Let 0 < a < 1 and let n be a positive integer. Show that as a and n > 00 such that na > p, the binomial probability (:) aX (1 atX
> 0
5 Distributions of Random Shapes
(4.102)
converges to the Poisson probability
p,x exp( _ p,)
(4.103)
x!
11. In Section 4.6.3, we noted that the space of undirected lines in the plane is homeomorphic to the Moebius strip. In this problem, we shall go through the steps to prove this fact. (a) Any undirected line can be directed in two possible ways. I~ one of those directed lines has coordinates (r,O), show that the other hne has coordinates (r, 0 + 'If), where summation of angles is performed modulo 2'1f.
(b) We can embed the cylinder of directed lines in R
3
{(x,y,z): y2 + z2 = I}
as (4.104)
which is homeomorphic to R x S1. In terms of these threedimensional coordinates, plot the points
{(r, cos(O), sin(O)), (r, cos(O + 'If),sin(O + 'If))}
(4.105)
and note that the line in R 3 that passes through these two points passes through the origin. (c) Argue that the space of undirected lines in the plane is homeomorphic to the space of lines constructed in part (b). This space is not the projective plane Rp 2 because there is a line through the origin missing. Which one is it? (d) From part (c) above, we conclude that the space of undirected lines is homeomorphic to the projective plane Rp 2 with one point removed. Use Problem 4 from Chapter 2 to argue that such a space is homeomorphic to the Moebius strip. (Hint: removing a point from a manifold is topologically equivalent to removing a closed disk. Why is this?)
5.1
Landmarks from the Spherical Normal: lID Case
We are now in a position to state and prove a central result, due to Kendall [90], for the induced distribution of the shapes of planar landmarks generated by an lID spherical normal model. Proposition 5.1.1. Let X 1,X2 , ••• ,Xn , n ~ 3, be independent and identically distributed spherical normal variables in R 2 • Let
(5.1)
be the shape representation of the points as an element of E~ ~ Then a has a uniform distribution on E~.
cpn2.
Proof. This result now follows directly from two results in Chapter 4. From the remarks at the end of Section 4.5.3, we note that the preshape T E s2n3 has a projected normal distribution. The density function for this distribution in given in formula (4.63) with v = O. From Problem 9 of Chapter 4, we conclude that this distribution is uniform on s2n3. The result then follows immediately by applying Section 4.4.3 using q = n  2. Q.E.D. As a special case, we note that the shape of a random triangle of spherical
150
5.
Distributions of Random Shapes
5.1 Landmarks from the SphericalNormal: IID Case
normal variables is uniformly distributed on the sphere 8 2(1/2). A simple geometric application is the following:
151
So the joint density of U, V, Zl, ..., Zn2 with respect to the volume element dV2(u)dV2(v) fI;,:i dV2(Zj) is n2
Corollary 5.1.2. Under the conditions ofProposition 5.1.1 above, for' n ';= 3 the probability that X lX2X3 forms an acute triangle is 1/4.
Iv1 2n  4
4
f(u  v)f(u + v)
II feu + VZj)
(5.7)
j=l
Proof. Let us use the representation of 8 2(1/2) in R 3 given in formula (3.6). We begin by calculating the probability that the angle at vertex X 3 is greater that 7f /2. This corresponds to the region on the sphere where WI > 1/4. However, from Problem 8 at the end of this chapter, we note that the surface area of a sphere cut off by parallel planes is proportional to the distance between the planes. Now, since the sphere 8 2(1/2) has unit diameter, it follows that the probability that the angle at X 3 is greater than 7f/2 is 1/4. Similar results hold for Xl and X 2 by symmetry. As a triangle can have at most one internal angle greater than 7f/2 it follows that X lX2X3 is obtuse with probability 3/4. The result then follows immediately. Q.E.D. As complex projective spaces are hard to visualize, it is useful to write out the density function in terms of Bookstein coordinates. Once again, the arithmetic of the complex plane is useful. Let n;:::: 3. We introduce a transformation of complex variables
(Xl, X 2, ... , X n)
~
(U, V, ZlJ ... , Zn2)
Next, we integrate over the variables u and v to get the density function for the Bookstein coordinates. This has general representation as 4
fc fc
n2 I
v1 2n 
4
feu  v)f(u + v)
V
+
X2
X2
Xl
and Zl, ..., Zn2 are the Bookstein coordinates of the shape of Xl, ..., X n . Let us suppose for the moment that the landmarks Xl, ..., X n are independent, and have absolutely continuous distributions in the complex plane C with common density f(x). The joint density of Xl, ..., X n is
(5.9) Let f~ be the density of "Zl' ... ,Zn2 in formula (5.8). Then tt
n tt
en  2)!
n2 JCnl(
ZlJ ... , Zn2
)
(5.10)
(5.11) where (5.12)
(5.5)
j=l
Under the change of variables in (5.2), the volume elements transform as n2 dV 2(u) dV2(V)
4
_
Proof. Without loss of generality, we can scale the distribution of the landmarks so that the covariance matrix of all Xj's is the identity matrix. Plugging the normal density into formula (5.8) we obtain
n
II [J(Xj) dV2(Xj)]
41v1 2n  4
(5.8)
be the density function for a spherical normal distribution in the complex plane C centered at the origin. We define
(5.2)
(5.4)
2
n
dV2(u ) dV2(V)
(5.3)
2
=
+ VZj)
Proposition 5.1.3. Let f
where
Xl
feu
When the density under consideration is spherical normal, then this iterated integral can be computed exactly as follows:
f (Zl' ... , Zn2) 
U
]1
II dV2(Zj)
j=l
Some simplification is obtained by expanding the absolute values in formula (5.12) using (5.13) We can also write
(5.6) (5.14)
152
5.
5.2 Shape Densities under Affine Transformations
Distributions of Random Shapes
is a linear transformation of the plane that is area preserving, so that ~ L For n ~ 3, let
Next, we complete the square in the exponent with respect to the variable u. After some reorganization, we find that (5.11) becomes
.Jh
r IvI2n4exp(_KlvI2) reXP(~luVl::>jI2) dV2(U)dV2(V) (27f)n lc 2 lc 2 n
(5.20)
_4_
be lID and continuous with density f. Now suppose that we let the landmarks Xl, ... , X n be jointly transformed by the common linear transformation h. Define
(5.15) The inner integral can be computed by changing variables, expressing uv 'L ZjIn in polar coordinates (p, B). In polar coordinates, we can express the volume element as dV2(u) = p dp dB. Computing the inner integral, we find that (5.H» reduces to
(5.21) for j = 1,2, ..., n. A simple transformation of variables argument shows that the density of Yj is fohl.Let Zl"",Zn2 be the Bookstein coordinates for the shape of the landmarks Xl, ... , X n , and let WI, ..., Wn  2 be the Bookstein coordinates for the shape of the transformed variables Y 1 , ... , Y n . We will represent the density of Zl, ..., Zn2 by
(5.16) Next, we change v to polar coordinates to compute the outer integral. After a routine integration over the two polar coordinates of v we obtain the formula given in (5.10). Q.E.D. For n to be
f"(Zl' Z2, ..., Zn2) and those of WI, ..., Wn 
2
(f
0
= 3 we obtain the density for the Bookstein coordinate Z = Zl 3
fi (z) =
7f(3 + Iz1 2) 2
153
(5.22)
by
h 1)"(W1' W2, ..., Wn2)
(5.23)
In this section, we consider how the shape densities f" and (f 0 h l)" are related. Replacing f by f 0 h l in formula (5.8), we see that (f 0 h 1 ) " has integral formula
(5.17)
When n = 4 the density of (Zl' Z2) reduces to (5.18)
4
JJivI2n 4f[u'  v']j[u' + v'] n2 II flu' + (VZj)'] dV2(u) dV2(v)
(5.24)
J=l
Formula (5.10) is a special case of the shape density derived by Mardia and Dryden [116], in which their parameter T (not to be confused with our notation for preshape) goes to infinity.
where u' = h 1(u), v' = hl(v), and (VZj)' = hl(vzj). For convenience, we recycle notation a bit, replacing u' by u and v' by v. Note that h has unit Jacobian. So the integral reduces to
n2
5.2
4
Shape Densities under Affine Transformations
JJIh(v)1 2n4f(u  v)f(u + v) IT f {u + hl[zjh(v)]} dV2(u) dV2(v) J=l
5.2.1
(5.25) Now let us write v in polar coordinates as (p, a). Writing the area element dV2 (v) as p dpda the integral expression becomes
Introduction
In this section we shall use formula (5.8) to study the transformation of shape densities when landmark variables are themselves transformed by an affine transformation of the plane. As shape distributions are unaffected by translations and scale changes, it is sufficient to study the effect on shape distributions of linear transformations of the plane that preserve area. We follow the development given in Small [155]. Suppose (5.19)
41 1L 2
00
Ih(V)1 2n4 f(uV)f(u+V)}1f {u
+ h 1[zjh(v)]} dV2(u) pdpda (5.26)
Suppose we now define
'1
1
(5.27)
154
5.
Distributions of Random Shapes
5.2 Shape Densities under Affine Transformations
155
5
Then our integral becomes
jlh(';')I'n' {4JjP >n ' f(u 
v)f(u + v)
II
f(u
+ vz;.ldV,(u)dp }
da
(5.28)
Let us now restrict the class of densities f under consideration to those that are circularly symmetric about the origin in e. By this we mean that the level curves of the density f are circles centered about the origin, or equivalently, that the distribution is invariant under rotations of the plane about the origin. Then exploiting this symmetry, we see that the expression {} in (5.28) is equal to 1 ~ 27rf (Zl'" , Z2"" ..., Z(n2)"')
(5.29)
4 3
.. 1
2
o 1
Thus we obtain the following proposition: Proposition 5.2.1. Let f be circularly symmetric about the origin in C. Then
(f
0
l h )U(Zl ' ..., Zn2) =
2~
1 2
11"
jh(ei "')12n 
4
f U(Zl"" ..., Z(n2)",) da (5.30)
where
Zj",
4
is as defined in formula (5.27).
It should be noted that formula (5.30) makes no reference to the evaluation of densities in the original space of landmarks en. So under the symmetry assumption, (fohl)u can be computed directly from f~ without reference to f.
3
2
1
o
2
3
4
5
FIGURE 5.1. The shape density for the elliptical normal. The shape density has been plotted in Bookstein coordinates for graphical convenience. The density displayed, however, is relative to the uniform probability distribution on the sphere of shapes E~.
Plugging formula (5.17) into (5.30) and grinding out the integral, we get
5.2.2
Shape Density for the Elliptical Normal Distribution To illustrate Proposition 5.2.1, consider the case where f is spherical normal and n =3. Suppose we,consider a linear transformation ~(Z)
+ i~(z)
+ 81/2~(z) + i81/2~(z)
(5.31)
that stretches the plane, taking the circle
1
(5.32)
into the ellipse (5.33)
The density function f with covariance matrix
0
h will be that of an elliptical normal distribution
r, where r ll = sl, r 22 = 8, and r 12 = r 21 = o.
(fohl)tt(z)
=
3(8 + Sl) (5.34) 27r (3 + Iz1 2 ) 2 {1 + 3[(8  s1)~(z)/(3 + IzI 2 ) j2 } 3/ 2
Figure 5.1 shows a contour plot of the ratio (f a h1)U/f U, using the stretch factor 8 = 2. The function is symmetrical about the real axis, is maximized on the axis, and is minimized above and below at Bookstein coordinates corresponding to equilateral triangles. The level curves are circles. These circles become a little easier to understand if plotted on the shape sphere 8 2(1/2). Using the coordinates offormula (3.6), we see that the level curves on the sphere are of the form W3
= constant
(5.35)
As the spherical normal induces a uniform distribution on 8 2(1/2) it follows that these curves are the level curves of the induced density from the
156
5.
5.2 Shape Densities under Affine Transformations
Distributions of Random Shapes
elliptical normal distribution. The interpretation is then clear: the density induced by the elliptical normal is uniformly squashed towards the great circle of collinearities corresponding to W3 = O. The density is maximized on the great circle W3 = 0 and minimized at W3 == ±1/2. The reader should note that formula (5.34) becomes very simple if we restrict ourselves to evaluating the shape density for aligned sets of triangles. These are those for which S3(z) = O. In such cases, the formula becomes (j
5.2.3
h 1)P(z) = (s \Sl) f"(z)
0
(5.36)
Broadbent Factors and Collinear Shapes
The simplification in (5.36) for aligned landmarks is not unique to the normal distribution nor to the case n = 3. Let us return to the general case of formula (5.30). A general simplification in formula (5.30) is introduced if we restrict attention to those shapes that correspond to aligned sets of landmarks. These are shapes whose Bookstein coordinates (Zl' Z2, ... , Zn2) are real, so that S3( Zj) = O. When Zj is real, then Zjcr = Zj for all a :s; a < 27f. Thus our formula (5.30) reduces to (joh
1)P(ZI,
... ,Zn_2)
=
[2~
1 2
"
~
1 2
+ iS3(z)
>
s'1/2m(z)
+ is 1/2S3(z)
2~
1
"
Ih(eicr W
n 4
da
2
2
.
2+2+3s2
(5.41)
8
The circular symmetry used to obtain formula (5.37) is stronger than necessary. The following proposition weakens the symmetry assumption used to derive the Broadbent factor. Proposition 5.2.2. Let m be the least common multiple of the integers 2k  4, with k = 3,4, ... , n. Let f be a density for planar distributions that is invariant with respect to rotations by 7f/ m about the origin. Suppose h is an areapreserving linear transformation of the plane. Then on the collinearity set where Zl, Z2, ... , Zn2 are real, we have
[2~
(joh 1)P(Zl,,,,,Zn_2) =
1
27f
Ih(eicr ) \2n 
4 
jP(ZI, ... ,Zn2)
=
Ln 
2
(S\SI)
Proof. To prove this, we return to formula (5.28), and let a(a) be the expression {}. Furthermore, let b(a) = Ih(eicr)1 2 n  4 . Let us write out a(a) and b(a) in trigonometric series in the variable a. The function b(a) is a trigonometric polynomial of the form 1 27f
1
27f
o
.
Ih(e,cr)1 2n 
4
da
+
n2
L C1j cos(2ja) j=l
+
n2
L C2j sin(2ja)
(5.43)
j=l
From the rotational symmetry, we see that a(a) has a trigonometric series for which the coefficients of cos(2ja) and sin(2ja) are zero for j = 1,2, ..., n  2. The result then follows from the orthogonality of the trigonometric terms. All terms in the integrated product 27f
1 o
a(a)b(a) da
(5.44)
(5.39)
vanish with the exception of the products of the leading constant terms in the series. Q.E.D.
(5.40)
Of course, in the limiting form as n > 00 this is simply the circular symmetry assumed earlier. However, for n = 3 the assumption is only that f is invariant under rotations by 7f/2, a much weaker assumption.
as in (5.31). The corresponding Broadbent factor reduces to 2
t: (S+Sl) = 3s
(5.42)
i cr " \h(e W n  4 da (5.38) 27f 0 is known as the Broadbent factor, named after Simon Broadbent, who proposed its use and calculated some approximate values in [33]. The interpretation of these factors is straightforward. If we suppose that landmarks are initially generated by some circularly symmetric distribution, then, broadly speaking, shapes of landmarks will also tend to be rounded. If the distribution is then stretched by a linear transformation, much as a circle is stretched into an ellipse, then we naturally expect shapes of landmarks to be correspondingly elongated. This means that the shape density on the region corresponding to aligned landmarks undergoes an increase by the Broadbent factor. It can easily be checked that the Broadbent factor is always greater than or equal to one. For example, consider again the linear transformation
h : m(z)
where 12 m is the mth order Legendre polynomial. In particular, the first order Legendre polynomial is the identity function. Thus when n = 3, formula (5.37) reduces to (5.36) for all densities f that satisfy the circular symmetry condition of Proposition 5.2.1. In addition, we have
Ih(eicr Wn 4 da] f P(zl"",Zn2) (5.37)
The factor
157
158
5. Distributions of Random Shapes
5.3 ·Toois for the Ley Hunter
159
5.3 Tools for the Ley Hunter To illustrate the methods developed in the last two sections, let us consider a statistical problem that provided some of the impetus for the development of the Kendall school of shape analysis. In 1925, Alfred Watkins published The Old Straight Track [177], which proposed the imaginative hypothesis that a variety of megalithic sites in Britain were, in fact, curiously aligned along tracks he called leys. Watkins was an amateur archeologist with a fascination for folklore and mysticism, and his writings drew deeply upon the latter. In addition to sites marked by standing stones and burial chambers, Watkins also included the locations of churches, river fords, and certain place names, on the assumption that although the presentday marker is relatively recent, the site was chosen for its importance as part of the system of ley lines. Watkins' hypothesis is not to be confused with the alignment hypotheses of Alexander Thorn and his investigation of megalithic sites as ancient observatories. The ley hypothesis is unlikely to be settled by statistical argument, because the validity of folklore is not subject to direct statistical analysis. Those who find the arguments from folklore convincing may consider the statistical arguments irrelevant. On the other hand, the hardened empiricist may dismiss the issue out of hand. However, statistical problems of this nature are commonplace in archeology and deserve consideration as a family of similar questions. In many cases, the presence of patterns in such data can be interpreted as the consequenceeither of design or of chance, the latter interpretation usually based "upon the large number of combinatorial possibilities that the data provide. The ley line hypothesis is a case in point. For example, consider the coordinates of the 52 megalithic monuments in Cornwall, England known as the Old Stones of Land's End. These coordinates are displayed in Figure 5.2. While there are indeed many megalithic sites that can be connected by straight lines to a high degree of precision, we would normally expect a reasonable number of nearly perfect alignments by chance among such a large number. For example, among 52 landmarks, there are 22,100 triangles that can be formed with vertices among the landmarks, and 270,725 quadrilaterals of landmarks. In standard stochastic models, the probability that three or four landmarks are approximately collinear is small. Nevertheless, balancing this is the large number of subsets of triangles and quadrilaterals that can be formed. So we would expect a reasonable number of such collinearities purely by chance. Among the megalithic data sets, the Old Stones of Land's End have received considerable attention. Broadbent [33] proposed a statistical study of the alignments among these 52 sites, which are plotted in Figure 5.2. The reader can find the data set in [33]. The 52 sites are scattered irregularly across Land's End. Alignments of the sites can be drawn through the points. However, it is difficult to tell a priori whether these alignments are
40 l1li
35
I

l1li
l1li
l1li
l1li l1li
l1li
l1li l1li
l1li l1li
30

l1li
l1li l1li
,pa
l1li
2S

l1li
l1lil1lil1lil1li
l1li
l1li
l1li
l1li
l1li
l1li
l1li
l1li
.. l1li
"
l1li l1li
l1lil1li
1III';rJ'
l1li
20
35
40
4S
so
FIGURE 5.2. The Old Stones of Land's End in Cornwall, England. The 52 plotted points are based upon measurements by John Michell, Chris HuttonSquire, and Pat Gadsby. The horizontal axis marks the coordinates of thestones in an eastwest direction and the vertical axis the coordinates from north to south.
160
5. Distributions of Random Shapes
5.3 Tools for the Ley Hunter
coincidental. The first requirement is a definition of approximate alignment of sites or landmarks. A variety of definitions are possible. These are summarized in [33] and examined for their strengths and weaknesses in testing the ley line hypothesis. Overall, the angular criterion, used in [33] and [95], provides the best guarantee of accepting configurations of landmarks that might be intentionally aligned. Following [33] and [95], we adopt such an angular criterion.
Definition 5.3.1. Three landmarks X 1,X2,X3 will be said to be aligned to within tolerance f if the maximum internal angle of the triangle with vertices at X 1X2X3 is ~ 7r  f radians. We shall also say that the triangle X 1X2X3 is eblunt when this condition is satisfied. As there are 52 such sites or landmarks, in a random scattering of 52 points in the plane the expected number of such eblunt triangles will be 22,100 times the probability that any given triangle X j X k X, is eblunt. For the purposes of the analysis that follows, we shall assume that e is sufficiently small that the approximations that follow are reasonable. This will involve discarding higherorder terms in E, which is acceptable provided f is about one degree and certainly less than 5 degrees. Such values would be realistic given the technology available to megalithic architects. Suppose we model the landmarks as having an elliptical normal distribution in the plane. Let Xl, X 2 , and X 3 be lID elliptical normal random vectors in C. We shall assume that the eccentricity of the distribution is governed by the stretch factor s as in Section 5.2.2. For E < 1f/2 the probability that X 1X2X3 is eblunt is three times the probability that the internal angle at X 3 is greater than tt  E. In Bookstein coordinates, the region of shapes where the triangle X 1X2X3 is eblunt at X 3 is a lens bounded by the circular arcs that meet the real axis at ±1 making an angle E. See Figure 5.3. For small f, these circular arcs can be approximated by the parabolas
~(z) = ± f[1  ~2(Z)] ,
1~(z)1
:::; 1
(5.45)
When Xj, j = 1,2,3, are IID elliptical normal with stretch factor s, the shape density for the triangle will be as given by formula (5.34). In particular, we are interested in this shape density close to the set of aligned triangles. Thus we may assume that ~(z) ~
0
(5.46)
~ ~(z)
(5.47)
and
Izl
1
161
+1
FIGURE 5.3. Lens of blunt triangles in Booksieiti coordinates. In Bookstein coordinates, the region of cblunt triangle shapes is the union of three sets, each set corresponding to a vertex at which the internal angle is :::: 1f  E. The set corresponding to ebluni: angles at X a is the lensshaped region in the middle of the figure. The wedgeshaped region on the left of the lens corresponds to triangles where the Xl is Eblunt. Similarly, the wedgeshaped region on the right corresponds to an cblunt angle at X 2 • If these three sets are plotted on the sphere of triangle shapes, they are seen to be congruent to each other and of equal probability under an IID model for X IX2Xa . Applying these approximations in (5.34), we obtain
3(S+S1) 21f[3
+ ~2(z)]2
(5.48)
as the approximate shape density close to alignment. Integrating (5.48) over the region between the parabolas in (5.45) gives the probability that X IX2X3 is eblunt at the vertex X3. This must be multiplied by three to allow for blunt angles at the other two vertices. So the probability that X 1X2X3 is eblunt is approximately
(s+ sl) (9 2
1fV3)f 31f
(5.49)
Note that this formula clearly breaks down when s is large, because the approximation to the probability becomes greater than one. As s > 00, the shape distribution becomes squashed down onto the real axis in Bookstein coordinates, and the density is no longer approximately constant over the lens in the imaginary coordinate. Fitting an elliptical normal distribution to the scatterplot in Figure 5.2 gives an estimate of s = 1.6612, Thus we would expect on average 164.8 triangles that are blunt to within a tolerance E of one degree. In fact, there are 142 such triangles in the data set, which is within chance variation. Silverman and Brown [153] have shown that under the null hypothesis that the points are lID and continuously distributed in the plane, the distribution of the number of eblunt triangles is approximately Poisson for small values of E. Thus the number observed is about 1.77 standard deviations below the estimated mean under the null hypothesis model. This analysis is preliminary at best. Several questions remain. Has the value of E been chosen appropriately? Does the normal model represent
162
5.
5.5 Landmarks from the Spherical Normal: NonIlff Case
Distributions of Random Shapes
a valid null hypothesis? Should we be looking at alignments of more than three points? Finally, is the criterion of ebluntness an appropriate one for searching for leys? We cannot take the time to give satisfactory answers here. However, each of these problems can be dealt with briefly. 1. The value of E can be treated as a nuisance parameter of the problem. This analysis leads to the pontogram technique of Kendall and Kendall [95]. 2. The normal model is not the only mechanism that can serve as a null hypothesis of random alignment. Uniform scatterings in rectangles have been investigated in [33] and. [95]. Uniform scatterings in ellipses have been investigated in [95], [154], and [155]. The expected number of alignments in the uniform elliptical model is less than the elliptical normal model. However the observed number of alignments is still not statistically significant. 3.' Alignments of four points can be investigated. However, the evidence for ley lines does not appear to be much more convincing in this case either. See [33] for some simulations. There are so few alignments of five or more points that it is difficult to draw conclusions of any statistical significance. 4. The use of the maximum angle of the triangle as a measure of alignment is only one of several ways of defining approximate alignment of points. An alternative definition is the strip definition, under which a set of points is aligned if it falls entirely within a strip of given width, the width defining the tolerance much as E did for the angular criterion. Again, we refer the reader to [33].
5.4
Independent Uniformly Distributed Landmarks
Another model for lID landmarks that has attracted some attention is that for which the landmarks are uniformly distributed in some bounded convex region of the plane. Let A be such a bounded convex region in the plane with positive area, and suppose that Xl, X2, ..., X n are lID uniform in A with n ~ 3. Then
1/V2(A)
f(x)
=
{
xEA (5.50)
o
For this case, we can rewrite formula (5.8) for fn as .) = 2 1 f n(Zl, ... , . Zn2 2n 4[V2(A)]n
J')rB X2 I
2n
Xl
1

4
where
B = {(Xl,X2) E A 2 : xl(l zj)/2 +x2(1 + zj)/2 E A for all j}
=0
(5.54) These Bookstein coordinates correspond to aligned sets of landmarks with X 3 , ... , X n falling on the line segment from Xl to X 2 • Under a permutation of the labels of the landmarks, any aligned set of landmarks can be written in this form. Thus it is possible to find the density function for shape of any aligned set or the approximate probability of an approximate alignment of landmarks. It is the latter that was useful in studying the Land's End data of the previous section. Thus the uniform model provides an alternative to the normal model used earlier. Problems 2 and 3 invite the reader to analyze the Land's End data using the uniform model of points scattered in an ellipse or rectangle. The integral in (5.54) is closely related to the wellknown Blaschke constants of the convex set A. See [147, pp. 4649]. For example, for a circular disk A of unit radius, we find that (5.54) becomes 1 (2n) 22n31rn2(n 1)(2n  1) n
(5.55)
from which the value of fn can be evaluated when ~(Zj) = o and 1 $ iR(Zj) $ +1 for i > 1,2, ..., n  2. Multiplying by a Broadbent factor gives us .cn  2 [(s + s1)/2] (2n) (5.56) 22n31rn2(n  1)(2n  1) n which provides the corresponding density for an ellipse with stretch factor s. See Problems 2 and 3 for an application to the collinearity calculation of the Old Stones of Land's End. For values of the shape density off the aligned region in Bookstein coordinates, the integral is harder to evaluate. The function fn is typically complicated but can be found in closed form. See Le [101, 102] for some excellent work on this difficult problem.
5.5
Landmarks from the Spherical Normal: NonlID Case
(5.52)
and 1 $ iR(Zj) $ +1 then
xl(1zj)/2 +x2(1+zj)/2
is a convex combination of Xl and X2 and lies on the line segment from Xl to X2. As A is assumed to be a convex set, this convex combination will then lie in A, and the indicator function will equalone. If this is true for all j = 1,2, ... , n  2, then the shape density fn reduces to
dV2(Xl ) dV2(X2)
(5.51)
Note that when ~(Zj)
163
(5.53)
The main intention of this section is to prove a beautiful result due to Mardia [114] based on the work of Mardia and Dryden [116, 117], that the density function for the shape of nonlID spherical normal landmarks in
164
5.
5.5
Distributions of Random Shapes
the plane has a particularly elegant form. We follow the derivation given by Goodall and Mardia [70]. First of all, we will need a lemma. Lemma 5.5.1. Let 'I3 k == 2. Then
10r
be the function defined by equation (4.62) with
2rr
(5.57)
'I3 [t cos(B)] dB
Proof. Returning to formula (4.63), and setting v == t and n = 2, we see that 2rr exp( _t 2 /2) 'I1[t cos(B)] = 21f (5.58)
1
because the density in (4.63) must integrate to one. Rearranging (5.58) we get
10i" cos(B) exp[t2 cos2 (B)/2][tcos(B )]
v'27f[exp(t t
dB
2/2)
 1] (5.59)
Landmarks from the Spherical Normal: NonIID Case
165
I: IIJ.Lj
 ,u112,
where uJL is the shape of the triangle J.L1JL2JL3 and ,6 = with ,u = (1/3) I: JLj'
The inner product used in this density formula is that of R3 with the sphere 8 2(1/ 2) embedded in R 3 as a sphere of radius 1/2 centered at the origin. Note that this formula differs slightly from that given in [70]. We have chosen to construct the density function on the shape space 8 2(1/2) rather than to renormalize the radius of the sphere to one. The result is that our ,6 is four times the concentration parameter used in [70]. Before we derive this formula, some observations need to be made. This density function is unimodal, with maximum value at U JL and minimum value at the antipodal point uI" If we think of 0' JL as a north pole then the density function is seen to be constant along lines of latitude on the sphere. The constant ,6 acts as a measure of concentration of the density about U I" High values of ,6 produce distributions that are closely concentrated about U JL while low values produce distributions that are more diffuse on the sphere. In fact, as JLl' JL2, J.L3 converge to some common point in R 2 the constant ,6 goes to zero and the density function converges to 1f l = 1/V2[8 2(1/2)]. This is the uniform distribution on 8 2 (1/ 2) and is seen to be in agreement with the result in Section 5.1.
Differentiating both sides with respect to t we obtain
1
2rr
cos3(B) exp[e cos2(B)/2][tcos(B)] dB
(5.60)
= v'27f t 3 [1  t 2 / 2 + t2 exp(t 2/2)  exp(e /2)] Using formulas (4.644.66) we see that 'I3 (t )
= (2 + t 2 ) + 'l/J(t)(3t + t 3 )
(5.61)
where 'l/J(t) = if?(t)/1>(t). Plugging (5.61) into the lefthand side of (5.57), we can expand the integral into four terms. Applying the identities in (5.60) and (5.61) gives the required result. Q.E.D. Now, suppose Xl, X 2 , X 3 are independent landmarks, with Xj having a normal distribution centered at mean point J.Lj E R 2 . Suppose also that the three landmarks have a common covariance matrix, which is some multiple of the identity matrix. Without loss of generality, we can scale the problem 2(1/2) be the so that this covariance matrix is the identity. Let 0' E 8 shape of the random triangle X 1X2X 3 as expressed in Kendall's shape space E~ ~ 8 2(1/2). Then we have the following proposition: Proposition 5.5.2. The density of dV2 (0') is given by 1fl(l+,6
2, it nevertheless provides an excellent mechanism for the simulation of such simplexes without resort to generating a Poisson process. As spherical normal landmarks are easy to simulate, an acceptance method can be used that first siTulates a n~rmal simplex and accepts this simplex with probability X P Imax(x P ). This will generate realizations of the shapes of PoissonDelaunay simplexes. It can be demonstrated that X is minimized when X = 11 y'p + 1. See Problem 7. For further analysis and comments on this simulation technique, see [93].
5.7 Notes The development of this chapter roughly follows the historical order in which the distribution theory for shapes was developed. The earliest work by David Kendall, Wilfrid Kendall, and Simon Broadbent concentrated on the IID uniform and normal models with a possible eccentricity parameter. The calculation of collinearity probabilities and Broadbent factors was developed by Small [154, 155]. The earliest reference on the uniformity of shape distributions under a spherical normal model would appear to be by Kendall [87], where the model under consideration was that of the shape distribution of a set of points in the plane that diffuse independently from a common starting point as Brownian motions. Applications of the distribution theory were typically archeometric in nature. The value of general shape distributions for biometric and morphometric applications was developed by Fred Bookstein, Kanti Mardia, Ian Dryden, Colin Goodall, and others. Bookstein proposed a landmark model with normally distributed landmarks in which the landmark variability about their respective means is small compared to the distances between landmark means. This leads to the socalled "tangent approximation" (to use David Kendall's terminology) in which the shape statistics, when expressed in Bookstein coordinates, have approximately a normal distribution. This follows from a Taylor expansion of the formula for Bookstein coordinates, in which the dominant term is a linear transformation of the original normal variables. The discovery that
171
shape variables under such circumstances can be approximately normal is reassuring, because there is a large literature on multivariate statistical analysis that can be tapped. Such models correspond to the circumstance in Proposition 5.5.2 where the concentration parameter (3 is large. The idea of using statistical techniques on shape variables that are commonly associated with multivariate normal theory also arises in allometry, where the logarithms of size variables are jointly plotted and analyzed for collinearities whose presence supports the growth allometry model of formulas (1.1) and (1.2).
5.8
Problems
1. In [39], Lewis Carroll (Charles Dodgson) proposed a number of mathematical "pillow problems" that Carroll claimed to have solved in bed. On the evening of January 20, 1884, he stayed up late to solve the following, which became number 58 on his list of pillow problems. Find the probability that three points chosen at random in the plane have an obtuse angle. The solution given proceeds thus: Let X IX2X3 be such a triangle. Without loss of generality, we can assume that X I X 2 is the longest side. Then X3 lies in the lune that is the intersection of the two circular disks having centers at Xl and X 2 respectively, and common radius IXI  X 2 1. The triangle will have an obtuse angle if and only if X 3 lies in the circular disk with center at the midpoint of X IX2 and radius IXI  X 2 1. The ratio of the area of this circular disk to the area of the lune is 'Tr/8
'Tr/3 
v'3/4
(5.89)
which is taken as the required probability. (a) Comment on this solution, discussing its assumptions and its validity. (b) Compare the solution with Corollary 5.1.2. Which is more convincing? (c) Find the probability of an obtuse angle for three independent points that are uniformly distributed on the boundary of a circle. For further reading on this interesting problem, see [39] and [135J. 2. Let Xl, X 2 , X 3 be IID uniformly distributed in an elliptical region with stretch factor s as in 4.4.2. Find the approximate probability that the triangle having Xl, X 2 , and X 3 as vertices is eblunt to within a tolerance € of one degree. This can be derived following the pattern of Section 5.3, applying formula (5.56) for the elliptical uniform shape density close to alignment instead of the elliptical normal shape density. 3. Compute the approximate probability that X 1X2 X 3 is eblunt to
172
5. Distributions of Random Shapes
within a tolerance of one degree as in Problem 2 above using a rectangle instead of an ellipse. Assume the sides of the rectangle are in proportion s: 1. How does this compare with Problem 2? 4. Find the density function for the distribution of shape in Bookstein coordinates for a triangle of three independent points uniformly distributed on the circle x~ + x~ = 1.
6 Some Examples of Shape Analysis
5. For a triangle in the plane with vertices X 1,X2,X3 find the Bookstein coordinate Z in terms of the three internal angles of the triangle. Use this to find the joint shape density for the internal angles of a PoissonDelaunay triangle using Corollary 5.6.2. 6. From Problem 5 above. Suppose that one of these three angles is chosen at random. Show that the distribution of this angle B has density
~[(7r 37r
B)cos(B) +sin(B)]sin(B)
(5.90)
7. From Section 5.6, prove that X of formula (5.80) is minimized when X = 1/';p+ 1. 8. Find a formula for the surface area of a sphere bounded between two parallel planes intersecting the sphere. Using the fact that this surface area is proportional to the distance between the planes, fill in the details of Corollary 5.1.2.
6.1
Introduction
In this chapter, we shall consider in greater detail some examples that were first mentioned in Chapter 1. While the Land's End data of Chapter 5 were accessible to analysis largely by shape theory alone, most spatial data sets contain scale and orientation information that should not be ignored. In many cases, a shape analysis is performed in order to find the relationship between size and shape. This is of interest in growth allometry, as was mentioned in Chapter 1. However 1 the relationship between size and shape is of interest more generally, as is evident in the dinosaur footprints example described in Section 6.2 below.
6.2
Mt. Tom Dinosaur Trackways
In this section, we continue the investigation through a size, orientation, and shape analysis of the dinosaur trackways of Section 1.4.2. See Figure 1.4. The collection of footprints at this site has undergone considerable deterioration due to weathering and vandalism, making precise statistics on the distribution of footprint dimensions impossible to collect, as Ostrom noted. However, it was possible to classify the footprints into three groups, with the larg~st being tentatively identified as Eubronies giganteus, the inter
174
6. Some Examples of Shape Analysis
mediate size prints being also tentatively identified either as Anchisauripus sillimani or as immature Eubrontes prints (the former being favored), and the smallest prints as Grallator cuneatus. The dinosaur Eubrontes was an early, mediumsized, tridactylic bipedal therapod, believed to be predatory in nature. On the other hand, Grallator was a small dinosaur of the same period that was bipedal and tridactylic, either a predatory therapod or a herbiverous ornithischian. It was found that Eubrontes prints varied in length from roughly 28 to 35 em., with the caveat that erosion makes precise determination of dimensions impossible. The Anchisauripus prints varied in length from 15 to 20 em. approximately, and the Grallator prints varied from 9 to 12 em. in length. The data set was recorded on 20 x 20 graph paper at a scale of 10 ft. to the inch, the site being divided into fivefoot squares for the purpose. Upon examination, the footprints were grouped into trackways, with some element of uncertainty in a number of cases, particularly where trackways cross. Uncertainty also arises in deciding whether two trackways along a common line were made by a single dinosaur or by two. For example, it is unclear whether footprint D should be grouped with trackway 15 or separately. Similarly, prints A, B, and C could be grouped with trackways 7, 9, and 11 respectively. For statistical purposes, it seems appropriate to analyze within and between trackway variation in size, orientation, and shape only on that subset of footprints that can be clearly classified. The loss of information by so doing is less problematic that the difficulty of . outlier contamination by including all prints. Thus trackways 5, 6, and 7 are somewhat confounded with each other. For the purposes of statistical analysis, we take the first three prints of trackway 5, the first four prints of trackway 6, and the first two prints of trackway 7. When grouped according to species, trackway 7 (counting footprint A :;IS a continuation of trackway 7), trackway 27, and trackway 28 were classified by Ostrom as belonging to Anchisauripus. Trackway 14 and trackway 18 were Classified as Grallator. All the rest were classified as trackways made by Eubrontes. The greatest uncertainty in this threefold classification is in trackway 13, consisting of a single isolated print, and trackway 7, consisting of two prints.
6.2.1
Orientation Analysis
We have already considered in broad terms some of the orientation information in the trackways and its relationship to possible gregarious behavior of Eubrontes. There are two types of orientation information within a trackway: the orientation of the footprint and the direction of the. trackway. Footprint orientation is typically compatible with the orientation of the trackway, and should be considered of importance in an orientation analysis in providing an ordering to the prints along a trackway. However, there is considerable damage to the footprints, making orientation ora footprint
6.2 Mt. Tom Dinosaur Trackways
175
FIGURE 6.1. Distribution of trackway orientations. The Mt. Tom dinosaur trackways can be individually oriented by taking a unit vector pointing in the direction from the first observable footprint in the trackway to the last observable footprint. Such a unit vector can be regarded as a point on the unit circle centered about the origin in the plane. Figure 6.1 shows the histogram ofthe scattering of trackway orientations as they are seen in Figure 1.4. It is evident that the vast majority of the tracks point in a westerly direction. These tracks are largely Eubrontes footprints.
difficult to determine in isolation from other footprints in the same trackway. Thus the trackway orientation would seem to be of greater importance in the analysis. The trackways are fairly straight, with the exception of number 17, which shows a slight but systematic curvature. We can encode the directions of the trackways by taking a vector from the initial footprint to the final footprint of the trackway and standardizing the vector to have length one. The exception to this definition is trackway 13, which contains only one footprint. The orientation of this trackway must be established roughly from the orientation of the footprint. The result is a directional data set that can be plotted on the unit circle. A histogram of the orientations can be seen in Figure 6.1. To estimate the overall direction of dinosaurs crossing the area, the directional median seems appropriate as it is less sensitive to large deviations in direction away from the overall trend, in this case to the west. If (h, O2 , ... , On, 0 ~ OJ < 27f, are the angles of a set of n directional vectors, then the median of the angles is that value 0 rninimiz
176
6.2 Mt. Tom Dinosaur Trackways
6. Some Examples of Shape Analysis
ing the sum of geodesic distances Lj d(8, 8j) around the circle, where d(8, 8j ) = min(18  8j \, 211" 18  Ojl). For our data set, the directional median is achieved on an interval of angles intermediate between trackway 4, where 84 = 3.1363, and trackway 5, where 85 = 3.146. (The fact that these trackways are also consecutive appears to be a coincidence of Ostrom's numbering system.) Averaging the angles over this interval provides a convenient summary of the median direction. The median works out to be 8:::0' 3.14, which is very close to due west, placing due north on the vertical axis of the coordinate system. (Considerable continental drift has occurred since the period when the trackways were formed. Therefore due west at present does not correspond to due west during the period when the footprints were preserved.) A total of 20 out of 28 trackways fall within a narrow 300 interval about the median direction.
6.2.2
a
N
I
Scale Analysis
There are two measures of size within a trackway that are relevant to the analysis of dinosaur locomotion. These are the footprint length and the stride length, the latter usually defined as the distance between successive footprints. Figure 6.2 shows a set of 28 boxplots of the sample distributions of stride lengths along the 28 trackways. The variation in stride lengths between trackways is most evident between trackways 14 and 18, classified as Grallator, and those trackways classified as Eubrontes, a much larger dinosaur. (The reader who wishes to see a comparison of Grallator and Eubrontes footprints is referred to page 128 of [133].) The differences in stride lengths can be interpreted as due to two factors, the first being the length of the dinosaur's legs from hip to foot and the second being the speed of the dinosaur. Footprint dimensions give us some indication of the size of the dinosaur, from which it is possible to estimate the speed that the dinosaur had when crossing the site. We noted the footprint dimensions above. These values are variable, even within a trackway, and so it seems safest to use species averages alone in the formulas. Alexander [I, 2] has proposed a formula for dinosaur speed based upon footprint length and stride length using Froude numbers. Froude numbers and the associated theory proposed by Alexander suggests that if two bipedal animals of similar shape have a size ratio of a: b in linear dimensions then their speeds will be in the ratio va:..jij. This suggests that the appropriate formula linking dinosaur speed to stride length and footprint length is of the form speed = c x (stride length)" x (footprint length)p+o.5
(6.1)
for appropriate constants c and p. A regression can be performed on modern bipedal species to fit the constants c and p. From this fit we can tentatively estimate dinosaur speed. Using Alexander's empirical fit to a
177
q H
H
to
d
a
dL:;;::::;:;c;:;,=::c':::::::::==:::::::=:====::c:==c::=~~~,..,,___.J 1234567 8910111213141516171819202122232425262728 FIGURE 6.2. Boxplots of stride lengths for dinosaur trackways. Distances are shown in meters. The trackways are ordered along the horizontal axis from 1 to 28. Each trackway has its stride length distribution displayed by a vertical boxplot that is constructed as follows: Each box appears as a thin dark rectangle with endpoints at the upper and lower quartiles of the distribution. The white horizontal strip inside each box marks the location of the median of the distribution. At each end of the box, dotted lines are drawn to the most extreme data value that lies within a distance of 1.5 times the interquartile range. Short braces mark the ends of these dotted lines.
178
6.
6.2
Some Examples of Shape Analysis
variety of modern species, we estimate the speed to be speed = 0.49 x (stride length)1.67 x (footprint length)1.l7
(6.2)
where stride length and footprint length are in meters and the speed is in meters per second. The reader should note that we follow the majority here in defining stride length to be the distance between successive footprints. Alexander's definition is the distance between successive footprints of the same leg. The formula has been adjusted accordingly. Inserting mean stride lengths and mean footprint lengths for the three species, we estimate that Eubrontes was crossing the site at a speed of about 8.1 kilometers per hour, which is a reasonable jogging pace. On the other hand, Grallator is estimated to have been traveling at a speed of 7.1 kilometers per hour, which is not much slower, despite the difference in sizes of the dinosaurs. The intermediatesized Anchisauripus is estimated to have been the fastest of the three. Estimates for its speed are unreliable here because the paucity of tracks is compounded with uncertainty of identification of track 7 as Anchisauripus. However, the large estimate for the speed of Anchisauripus is due in great part to the large stride length of trackway 28 passing though the site in a northeasterly direction. Based upon equation (6.2) we can estimate the speed of this individual to be about 35 kilometers per hour. With such estimates, it would be very useful to be able to provide an error analysis. However, there are far too many systematic errors to regard these values as anything more than a rough indication. Not the least of the systematic errors is the necessity of using (6.2), which is based upon modern species, to describe dinosaurs.
6.2.3 Shape Analysis If all size variables were to scale in a similar manner, then we might expect two dinosaurs whose stride lengths were in the ratio a: b to have a similar ratio in leg length, footprint length, and speed. However, this is not the case for modern species and was almost certainly not the case for. dinosaurs. Differences in scaling will normally be reflected in differences in shape. So it is shape variation, and in particular the relation of size to shape, that becomes of interest. As mentioned in the scale analysis, two factors that influence stride length along a trackway are the size of the dinosaur and its speed. If we compare dinosaurs with similar Froude numbers, we find that the speeds of the individuals will be proportional to the square root of the size of the individual, assuming that Alexander's model for dinosaur speed is correct. This can noted from the exponents of formula (6.1). Thus, if we could group individuals with common Froude numbers together, a size variable such as leg length, for example, would be a scale variable for trackway dimensions. Within such groups, trackway shape distributions would be common, with
Stride
Mt. Tom Dinosaur Trackways
16
r,
i.s
f
179
.. .. . . L1~.11".:. ... ., . •
'4
f •
I~
l,
1 L
•
0.9
~ ••
0.8
f
if'
"' • •
0.1 0.6
~
o
__'_
____''__
0.1
O~
__'__ _.....J
0.'
Geodesic Distance
FIGURE 6.3. Plot of geodesic distance versus stride length for Eubrontes triangles. Along each trackway, three successive footprints form a triangle. Some of the size and shape characteristics of these triangles are plotted above. Horizontally, a shape statistic is plotted that computes how close the triangle is to collinearity. This is measured as the geodesic distance in the sphere 8 2(1/2) oftrianqle shapes from the great circle of collinear triangles to each shape point. Thus points plotted on the lefthand side of the figure represent nearly collinear triangles. On the vertical axis, the mean stride length of the three successive footprints is plotted in meters. The mean stride length can be defined as the average of two stride lengths: that from the first to second footprint, and that from the second to the third.
different trackways having different scale factors. However, in actuality, a dinosaur would have had considerable variation in speed much as modern animals do. This extra variation in speed beyond scaling effects for dinosaur size would be expected to appear as a "stretching" effect along the trackway direction. Thus if we consider the shape of the triangle formed by three successive footprints, the effect of increased speed along the trackway would be to stretch the triangle closer toward the great circle of collinear triangles in shape space 8 2 (1/ 2). Figure 6.3 illustrates this idea. The plotted points are statistics drawn from all triangles formed by taking three successive prints along Eubrontes trackways. On the horizontal axis is plotted the geodesic distance from the triangle shape to the great circle of collinearities (proportional to the absolute "latitude" taking the great circle of collinearities as the equator). On the vertical axis is the average of the two stride lengths of the footprint triangle, from first to second print and from second to third. As can be seen, the triangles at greater geodesic distances, which are closer to equilateral in shape, have smaller stride lengths. As the stride length is proportional
180
6.2 Mt. Tom Dinosaur Trackways
6. Some Examples of Shape Analysis
to estimated speed in our model, this means that triangles for which the estimated speed is greater tend to be flatter, as one would expect.
6.2.4
Fitting the MardiaDryden Density
To study the shape distribution in greater detail, we can fit a distribution such as the MardiaDryden density of formula (5.62) to the triangle shape data. One way to fit such a density to the data is by matching the centroid, or center of mass, of the MardiaDryden distribution to the centroid of the data. We recall that the shape space 8 2 (1/ 2) is naturally embedded in R 3 as a sphere of radius 1/2 centered at the origin. See formula (3.6). Thus the shapes of triangles of three successive footprints can be represented as points in R 3 . If 01, 02, ... , Om are a set of triangle shapes represented as points in R 3 , then the centroid of these points will be 1 m ii =

m
L:Oj
(6.3)
j=l
which will lie within the interior of the sphere 8 2 (1/ 2). The vector
(6.4)
211 ii ll
lies once again on the sphere 8 2 (1/ 2) and can be interpreted as the mean shape of the data. This mean shape plays much the same role for the data that the shape plays in the MardiaDryden density of formula (5.62). Correspondingly, the quantity lliill provides a measure of how concentrated the data are about the mean shape. Its analog is not [J of formula (5.62) as such, but rather
0,..
parameter analogous to and have level curves for the density that are circles of points that are equidistant from the location parameter. They also include a concentration parameter that is analogous to [J. For the dinosaur trackways, there are two shape distributions associated with sets of three successive footprints, namely those with two left prints and one right, and those with two right prints and one left. For the modeling of triangles of successive footprints, we can pool the information by assuming bilateral symmetry and reflecting alternate triangle shapes along a trackway. This reflection is equivalent to multiplying W3 by 1 in the coordinate system of formula (3.6). For many trackways, the confounding of prints with other trackways and the difficulty of distinguishing right and left prints still makes the task of pooling information from the two types of triangles problematic. However, for some trackways such as trackway 1, clear information seems to be available. In such cases, we can take triangles formed by two left prints and one right print as canonical, and reflect the shape distribution for every second triangle of three successive prints. To encode the shapes of triangles, we take the first and third prints of any sequence of successive footprints as the base of the triangle for Bookstein coordinates. Thus in the notation of formulas (3.1) and (3.2), the point X3 is in fact our middle footprint of the three. From the Bookstein coordinates we can encode the shape 0 of each triangle using the spherical representation (Wl,W2,W3) offormula (3.6). For trackway 1, the tabulated shape data are as follows: Triangle
Pattern
Wi
W2
W3
1
RLR
0.48647
0.02354
0.11310
2
LRL
0.48641
0.00490
0.11569
(6.5)
3
RLR
0.49371
0.00500
0.07889
The method of moments fit of the MardiaDryden density to a sample of triangle shapes is found by solving the equations
4
LRL
0.49909
0.03003
0.00316
5
RLR
0.49672
0.03147
0.04777
6
LRL
0.46722
0.02082
0.17684
7
RLR
0.47245
0.03951
0.15882
8
LRL
0.48762
0.07167
0.08422
9
RLR
0.49811
0.03717
0.02256
0,..
11£(0)11 = Iliill
0,...
(6.6)
in the unknowns [J and This fitting technique is closely related to the fitting of spherical data by the Fisher distribution using maximum likelihood estimation. As Mardia and others have noted, the MardiaDryden density is one of a variety of densities on the sphere, including the Fisher density, the projected normal density, and the Brownian motion density, which while functionally different, form flexible families of densities that are very close to each other in shape. Like the MardiaDryden density, these families have a location
181
182
. 6. Some Examples of Shape Analysis
6.3 Shape Analysis of Post Mold Data
183
It can be seen that the sign of W3 alternates in the table. So, we reflect the triangles of the RLR type. Fitting this to the MardiaDryden density gives us an estimate for D"p. and for{3, namely D"p.
= (0.49186,0.00297, 0.08980)
{3 = 452.5
(6.7)
The high value of {3 is indicative of the regularity of the footprint pattern, and supports the interpretation of Figure 6.2 that the Eubrotiies of trackway 1 was moving at a fairly constant speed. The estimated mean shape D"p. can be interpreted by noting its proximity and relationship to the shape (0.5,0,0), which marks a triangle of three equally spaced collinear points.
... . . . :.. : .. ... : .. ..:,.. . ..,.. ... . . .. " . 0'·
6.3 6.3.1
Shape Analysis of Post Mold Data
o
A Few General Remarks
In this section, we apply some methods of shape analysis to the post mold data that we first considered in Section 1.4.3. In particular, we shall examine statistically the interpreted roundhouses of Aldermaston Wharf and South Lodge Camp as shown in Figures 1.5 and 1.6 respectively. In 1973, an exhibition at the Institute of Contemporary Arts in London was entitled "Illusion in Art and Nature." It is interesting to note that one of the exhibits was a plan of an excavated Bronze Age settlement from Thorny Down in Wiltshire, England. The exhibit challenged people to interpret the configuration of post molds found at the site and to group them into recognizable patterns that would correspond to the original buildings on the site. See Gregory and Gombrich [76] for a discussion of the ambiguity of interpretation of this site in the context of the exhibit. The reader is invited to examine Figure 6.4, which.shows the layout of accepted post molds based upon J.F.S. Stone's excavation from 1937 to 1939. This figure should be studied in the light, say, of R. Wainwright's comment in A Guide to the Prehistoric Remains in Britain [175, p. 199] that the site contained the remains of nine circular houses called roundhouses. One strong indication of a roundhouse is in evidence. In other places, rough circles can be seen. However, these can equally well be interpreted as parts of structures that could conceivably have been rectangular rather than circular. The interpretation of nine buildings on the site follows directly from Stone's original report, which grouped the post molds into nine clusters. Stone found other evidence at the site such as the location of cooking holes and some pottery. Nevertheless, the post mold configuration provides most of the evidence for the number, location, and shape of the buildings. The archeological interpretation of such sites is assisted by a certain amount of background knowledge of the cultures that were present. Thus Thorny Down has been interpreted in the light of prior knowledge that
~.
e
. '.
r;
•
e.
. . e. .'
... : .
. . • e
.
o
.
o
o
e.
r
:% •
....". . ..'.
0
e
•
e
...".:: '0' :
''
~
. ...•
:
~ .:.~
•
_ _ _ _ _ _ _:501ecl
FIGURE 6.4. Post mold arrangement at Thorny Doum, Wiltshire. Thorny Down has achieved a certain amount of notoriety among archeologists for the ambiguity of its post mold pattern. In this picture, a large number of features, such as pits and cooking holes, have been eliminated so that the reader can judge the post mold evidence by itself It has been claimed that there were nine roundhouses on the site. One is clearly evident [romthe picture. This figure has been redrawn from
[165].
184
6.3 Shape Analysis of Post Mold Data
6. Some Examples of Shape Analysis
Late Bronze Age peoples of Britain typically built roundhouses with posts that were spaced 1.62.2 meters apart. However, such knowledge, while of great assistance, can be misleading. For example, in the case of the post mold evidence at Thorny Down, such background knowledge can lead the researcher to interpret circular buildings in cases where the interpretation is weak. The eye is very good at interpreting patterns in chaotic pictures but is not always reliable in its interpretations.
6.3.2
The Number of Patterns in a Poisson Process
Suppose the researcher observes a point process, such as a post mold pattern, within a twodimensional region. After studying the configuration of points, the researcher comes to believe that rather than being random, the particles of the process exhibit geometric regularity that cannot be explained by chance. For example, the particles could be arranged roughly in rectangles or circles, or perhaps have an approximate lattice structure. A null hypothesis that no structure exists, so that the perceived configurations arise by chance, could be modeled by a Poisson process or any other standard model for particles in the plane. Then the number of patterned configurations observed in the data can be compared with the expected number obtained by chance under the null hypothesis. We have already encountered one such example based upon configurations of straight lines when we studied the hypothesis of ley lines in the Land's End data. We now broaden the question to include the kinds of configurations that can appear in post mold interpretations. Suppose that X I,X2 , ... ,Xn are n random planar points for n ~ 3. For convenience, let us write X = (Xl, ... , X n). Similarly, we will write x = (Xl, ... ,xn ) for any realization of X. Let ((X) be a function of these points taking values in the set {0,1}. We can think of ((X) as an acceptance function, which notes that X has a certain property by assigning the value ((X) = 1 when X has the property and ((X) = 0 when it does not. Now let us further suppose that ((X) is a function of these points only through their shape so that
The function w is typically a measure of the size of the configuration. As with the function (, we shall suppose that w is a symmetric function of its arguments. Now suppose we observe a Poisson process in the plane throughout some planar region A, and decide to count the number of configurations X of n particles for which ((X) = 1 and Wo ~ w(X) ~ WI. Some configurations that satisfy these conditions will lie entirely within the region A and will be observed, while other configurations outside the window will not be observed. Configurations that overlap, having some points within and some without, will not be observed. Let N be the number of configurations X of n particles observed within A such that ((X) = 1. We shall be interested in finding the approximate distribution of N and making a comparison of this distribution with an actual count of configurations in a post mold pattern. Let aA denote the boundary of A. The exclusion of configurations that overlap 8A is understood as a boundary effect which is vanishingly small as A expands to fill the entire plane. Suppose the Poisson process has intensity p. Then N has expectation of the form given by the following proposition: Proposition 6.3.1. There exists a constant c() depending upon the choice of shape function ( such that the expected value of the count statistic N is c(N) = c(() p" (wi n2  w~n2) V2(A) (6.10) if boundary effects are ignored. Proof. For convenience in this proof, we shall assume that w(x) is the diameter of z: We perform a transformation of variables from
where w = w(x)j the angle () E S is the direction of the vector X2  Xl, defined except when Xl = X2; and a E ~2 is the shape of Xl, ... , X n . Then we can factorize the geometric measure on x as
(6.8)
for any complex numbers a, b with a oF O. As we are not particularly interested in the labels of the points, but rather in their geometrical characteristics as a point set, we shall also suppose that ( is a symmetric function of its arguments. Next, we suppose that w(X) is a nonnegative real valued function that is invariant under translations and rotations, and homogeneous under scale changes of the points. That is, w has the property that
w(aX I
+ b,
... , aXn + b) =
lal
w(X I ,
... ,
X n)
(6.9)
185
(6.12) For the general theory of such factorizations, the reader is referred to [4, 5J. When Xl, ... , X n are IID uniform in A then X is uniformly distributed in An = A X ... X A. Let W = w(X). Then
c[((X)I(wo:sw,,;wdJ =
[V2(~)]n
in ((x)l(wo~w:swd
dV2n(x)
(6.13)
Applying the factorization of formula (6.12), integrating over the variables and Xl, and ignoring the boundary effects of configurations X that
W, (),
186
6. Some Examples of Shape Analysis
lie within a distance of
WI.
6.3
from 8A, we see that the expectation becomes
We shall continue to ignore boundary effects in subsequent formulas. Now suppose that m > n particles are uniformly and independently scattered throughout A. Then the expected number of configurations X of n particles among the m that satisfy the shape condition ((X) = 1 and size condition wo:::; w(X) :::; WI. is (6.15) We take a Poisson limit by letting m . 00 and letting A expand to fill the plane so that m/V2(A) . p and VI.(8A)/V2(A) . O. Then the expected number of configurations of nparticles is seen to be asymptotically (6.16) We get the required formula by setting
Shape Analysis of Post Mold Data
187
The second comment to be made on this result is the presence of some potentially high exponents informula (6.10). When modeling particle scatterings as Poisson processes, we typically have to estimate the intensity p. Now, if we were investigating the presence of roundhouses of eight posts, say, and were to underestimate the intensity of the scattering by twenty percent, then we would underestimate the expected number of circular configurations that could be explained as chance by a factor about 0.168. One way to underestimate the value of p is to assume that the region A is the region of excavation as marked by the bold line in Figure 6.4. The natural estimate for .p is then the average number of post molds per square meter across the region of excavation. An examination of Figure 6.4 shows that the post molds are not homogeneously scattered across the region of excavation. An improvement on this assumption is to suppose that the post molds are homogeneously scattered across some subregion of the region of excavation that we can call the region of post mold activity. This may explain why simulation studies such as those described in [44] have found a larger number of circles at Thorny Down than can be expected from a Poisson scatter over the region of excavation. Unlike the region of excavation, the region of post mold activity is unknown. Therefore we cannot directly find its area as a way of estimating p. Fortunately, other techniques are available to estimate p, One method is to fit the theoretical distribution of distance from a typical post mold to its nearest neighbor.
(6.17)
6.3.3 An Annular Coverage Criterion for Post Molds which completes the proof. Q.E.D. We usually wish to know more than simply the expected value of N. The Poisson approximations of Silverman and Brown [153] are useful to determine this. They show that under certain asymptotic conditions described in [153] the distribution of N is from the Poisson family. As the expected value can be approximated by Proposition 6.3.1 above, the approximate distribution of N can be specified. These asymptotics appear to be quite reasonable for post mold investigations. Two comments on Proposition 6.3.1 should be made. As mentioned 'above, formula (6.12) is a special case of a factorization calculus in which geometric measures can be shown to factorize in location, scale, orientation, and shape components. See [3,4, 5] for work by Ambartzumian and colleagues on these factorizations. Note that we can pull out a shape measure from the factorization. This corresponds to our function f in (6.12) above. The resulting shape measure might be thought of as canonical. However, upon closer examination, it is seen to depend upon the choice of size function. As we have observed on previous occasions, shape constructions cannot be fully separated from the definitions of size variables that are used in the standardization of data sets.
An acceptance criterion that has been popular among archeologists studying and simulating post mold patterns is the coverage criterion. For example, Cogbill [44] and Litton and Restorick[109] have searched for patterns in post mold data by moving a set across the region of excavation. Cogbill searched for roundhouses by running an annulus, or circular ring, with fixed inner and outer radius over the post mold points. Annuli consisting of all points y E R 2 such that (6.18) were used, where € > a and w > a are constants controlling the thickness and inner radius of the annulus, respectively. The constant a E R 2 controls the location of the annulus, and was allowed to vary as the annulus was shifted over the region of post molds. Any set of n post molds that could be covered by an annulus for some choice of a would be declared an acceptable configuration and considered as a potential roundhouse. See Figure 6.5. In view of Proposition 6.3.1 we naturally seek an approximation to the distribution of the number of sets of n post molds that satisfy the annular criterion of Cogbill under the hypothesis that the post molds are scattered
188
6.3 Shape Analysis of Post Mold Data
Some Examples of Shape Analysis
6.
• • • I
I
, "
.
'
I
:
I I
I
I ,
, I
"
,
..........
__
.....
•
I
I
\
4 7r w 2 Q2(2) ~ V2(A)
,
I
,
•
\
'
\
",
,
'
\
\
•
I
I
I , \
,, ,
evaluated by calculating Q2(2),..., Qp(p) directly. For our particular application, we have p = 2 and a family of annuli given by formula (6.18). It is easy to check that
•
  ,
,
.1
I
I
;
; ;
,
"
'
•
Qn(n)
_ ...
•
••
FIGURE 6.5. The annular criterion for accepting a configuration of post molds. Cogbill (44J and others have proposed the detection of circular post mold configurations by running an annulus across the region where post molds occur. If a sufficiently large number of post molds can be covered in a given position of the annulus, these post molds are accepted as a possible roundhouse.
as a Poisson process. We need only find the expected number of circular configurations under the annular criterion and then apply the Poisson limit theorem of Silverman and Brown. Suppose C is some subset of the plane R 2 . We represent the translates of C as C(a) = {y + a : y E C} (6.19)
for each a E R2. Let X I,X2,,,,,Xn be IID uniformly distributed in the region A. What is the probability that there exists a translate C(a) such that X j E C(a) simultaneously for all j = 1, ... , n? That is, what is the probability that C can be translated to completely cover all the points Xl, ..., X n ? The solution to this problem will allow us to find the analogous expectation for Poisson processes. Mack [111] has solved this problem, not only for subsets of the plane, but also for the generaldimensional problem. Using the terminology of [111] we let Qn(n) be the probability that n independent uniformly distributed particles in A c RP can be covered by a translate of a given subset C c RP. We assume that Vp(C)« Vp(A) and that boundary effects of A are ignored. Then Qn (n) has the general form Qn(n)
=n
pl
[
for small E > O. Of course Ql(1) in formula (6.13) we obtain
1 + I)j(n  1)(n  2) ...(n  j) j=l
for some choice of the constants bl , b2 , ... , bp 
l.
]
[Vp_((C)] Vp A)
nl
(6.20)
These constants can be
~
n [E
+ (n
(6.21)
= 1. Solving for
,
I
;
189
 1)] (27r)n1
bl byevaluating Q2(2)
2] nl [V~A)
(6.22)
again for small E > O. We can check this formula by calculating Q3(3) directly without appeal to formula (6.22). Using the transformations of Section 1.2.3 of [147, pp. 1617] and integrating over formula (2.18) of [147] we can show that the probability the radius of the circumcircle through X I,X2,X3 is less than or equal to some value w is 67r 2W4/[V2(A )]2 for large regions A, ignoring the boundary effects. For small E > 0, our probability Q3(3) is approximately the probability that the radius of this circumcircle is between wand w(l + E). This reduces to Q3(3) ~
24 7r 2 w 4 V (A )2
E
(6.23)
2
Formula (6.23) can be seen to agree with (6.22) for n = 3 to first order in E> O. We are now in a position to write out the formula for circles in a Poisson process. Proposition 6.3.2. In a Poisson process of intensity p within a window A the expected number E:(N) of circular arrangements of n 2: 2 particles under an annular coverage criterion with annuli of the form given in (6.18) is £(N) ~ (21r)nl pn W 2n 2 En 2 V2(A) (6.24) (n  2)! to lowest order approximation in
E> O.
Proof. This is a straightforward consequence of (6.15), taking a Poisson limit as A expands to fill the plane and the number of particles m goes to infinity so that m/V2(A) + p. Q.E.D. The similarity between the formulas of Propositions 6.3.1 and 6.3.2 can be seen. For small E > 0, the annular coverage criterion factorizes into shape and size criteria that relate Proposition 6.3.2 to 6.3.1. Note also that the Poisson limiting distribution for N holds here as well.
190
We are now in a position to try such methods on post mold data sets. While the post molds at Thorny Down represent one of the most famous examples of ambiguous interpretations, the simplicity of the configurations at Aldermaston Wharf in Figure 1.5 and South Lodge Camp in Figure 1.6 of Section 1.4.3 make them better starting points for analysis.
6.4
Case Studies: Aldermaston Wharf and South Lodge Camp
Before it was discovered, Aldermaston Wharf was heavily plowed. Thus it can be reasonably assumed that some of the post mold evidence was destroyed by plowing. As the evidence is quite fragmentary, it is necessary to assess the strength of the interpreted roundhouses as carefully as possible. See Figure 1.4. The shaded regions are features of the site that are later than the time of the Late Bronze Age settlement..The irregular unshaded regions represent pits at the site. The archeological report on Aldermaston Wharf can be found in Bradley and Fulford [30]. The site at South Lodge Camp, shown in Figure 1.5, was reexcavated, and reported by Barrett et al. [9]. A number of buildings were identified and labeled A through D, with varying degrees of geometric regularity in the post mold evidence.
6.4.1
6.4 Case Studies: Aldermaston Wharf and South Lodge Camp
6. Some Examples of Shape Analysis
Scale Analysis
For a Poisson process of intensity p, the median distance from any particle to its nearest neighbor is J(ln 2)j(rrp). This suggests that we estimate p at Aldermaston Wharf by computing the median nearest neighbor distance and solving for p. The median distance from any post mold to its nearest neighbor is 1. 7 meters. So the intensity of the post mold scattering at Aldermaston Wharfis estimated to be P.= 0.076 post molds per square meter. A total of m = 61 post molds are scattered throughout the region, suggesting an area of post mold activity of mj P = 802.6 square meters. This is considerably less than the area of excavation, which was approximately 2000 square meters. From observations at other sites, roughly contemporary with Aldermaston Wharf and South Lodge Camp, we would expect neighboring posts of buildings to be within 1.6 to 2.2 meters of each other. Counting replacement posts and the exceptional larger distance, we would expect posts belonging to a common building to be less than three meters apart. Figure 1.4 shows all post molds that satisfy this joined by a link. Neither interpreted building I nor interpreted building II is clearly defined by this linkage method. However, rough circles can be made out for both I and II, with the strength of the circular interpretation being somewhat vague. We shall examine these circles more carefully below in
191
the shape analysis. At South Lodge Camp, a total of m = 71 post molds were recorded across the area of excavation, and the median nearest neighbor distance was found to be 1.4 meters. Using the same procedure for estimating p as was used for Aldermaston Wharf, we obtain an estimate p = 0.112 post molds per square meter for a Poisson process with the same median nearest neighbor distance. In turn, this allows us to estimate the area of the region of post mold activity to be mj p = 634 square meters. Again, this is considerably less than the area of excavation, which is about 1600 square meters. Linking post molds that are within three meters of each other we see that interpreted structures B, C, and D become clearly defined, with circles evident in D and C. Structure A is less clearly defined by this linking method.
6.4.2
Shape Analysis
The report on Aldermaston Wharf by Bradley and Fulford [30] interpreted two roundhouses, which are labeled I and II in Figure 1.4. Of the two interpreted structures, the second has the stronger visual evidence of a circular configuration. A total of 6 to 8 post molds can be interpreted as possible locations for posts of a roundhouse wall. There is some evidence that on the east side of Structure II a post mold could be missing because of the presence of later features. If this is the case, an interpretation with 6 posts as in Figure 6.6 would be appropriate. We assess the fit to a circle by choosing annuli that cover the configurations having the smallest possible area. Structure II can be covered by an annulus whose inner radius is 3.66 meters and whose outer radius is 3.95 meters. According to formula (6.24) the expected number of configurations of six particles in a Poisson process of intensity p = 0.076 throughout an area of 803 square meters is 1.07. Thus such a circular configuration can be considered plausible on chance considerations alone. Structure I is cruder than Structure II with an even higher value for e. The six post molds of Structure I shown in Figure 17 can be covered by an annulus with inner radius 3.15 meters and outer radius 3.62 meters. The expected number of such configurations over 803 meters for an equivalent Poisson process is 3.03. At South Lodge Camp, Cluster D contains a circle of eight post molds as in Figure 6.6, and can be covered by an annulus of inner radius 3.95 meters and outer radius 4.21 meters. In a Poisson process of intensity p = 0.112 the number of circular configurations of eight particles that can be covered by such an annulus within a region of area 634 square meters is 0.15. Thus the circular configuration of Cluster D is more unlikely than either Structure I or Structure II at Aldermaston Wharf. Cluster C also contains a circle of seven post molds. This can be fit by an annulus with inner radius of 2.13 meters and an outer radius of 2.36 meters. The expected number of such configurations is 0.00096. The reader may find it a bit
192
6.5
6. Some Examples of Shape Analysis
Automated Homology
193
unusual that configurations with the fit of Cluster C are much more rare than those of Cluster D in a Poisson process, as the eye suggests otherwise. The eye also picks up the symmetry of spacing of the post molds in Cluster D as a component of the regularity. The main difference that explains the calculations is that Cluster C has a circle of smaller radius. An examination of formulas (6.10) and (6.24) shows that small configurations are much rarer than large ones for a Poisson process. A hint of a circle can be seen in Cluster A of South Lodge Camp. However, the configuration is very weak.
6.4.3
The case studies and formulas have not provided a clear decision procedure that would allow as to accept or reject an interpretation of a roundhouse in a post mold data set. However, they do provide us with a quantitative tool for assessing the strength of a circular configuration relative to other configurations at the same site and relative to configurations at other sites. It is perhaps the latter that is more important. The interpretation at some famous sites such as Thorny Down is problematic, whereas the interpretation at sites such as South Lodge Camp is much more straightforward. Archeologists who can supplement their post mold analysis by comparing it quantitatively with other sites can evaluate the strength of their conclusions in the context of what is known about Late Bronze Age sites. For example, we can conclude that the interpretation at Aldermaston Wharf is tentative at best, with an interpretation that is weaker than that for South Lodge Camp.
a
•a South Lodge Camp Cluster D
o
•
o •
0
•
•
Aldermnston Whnrf
•
• • •• Structure II
•
Structure I
0 0
•
•
•
Conclusions
• • •
6.5 o
o
FIGURE 6.6. Roundhouse Interpretations at South Lodge Camp and Aldermaston Wharf
6.5.1
Automated Homology Introduction
In this section, we shall describe an automated homology routine developed by Michael Lewis as part of his Ph.D. work at the University of Waterloo. Up until this point we have represented various shapes by assuming that they are naturally homologous or that homologous landmarks can be selected from the data, as in the case of the brooch images of Chapter 1. However, in many image data sets there are no obvious features that stand out from the rest of the image to the extent that we would wish to label them as landmarks. We would rather seek to find a mapping from each image to any other that maps each point on the image to its homologous point on the other image. The problem of constructing a homology between images is closedly connected to the correspondence problem in computer vision, in which one has two images, each in two dimensions, of a threedimensional object seen from two angles. Ifit is known which points in the two images are different views
194
6. Some Examples of Shape Analysis
of the same point in three dimensions" the images are said to be registered. If there is no aspect of the object that.is visible in one image but not in the other, then the points in the images are homologous, with corresponding points between images being homologous if they are projections of the same point of the threedimensional object. The images will differ in shape slightly because of the two aspects from which the object is viewed. The main difference between the correspondence problem of computer vision and the automated homology problem of shape analysis is that the shape differences of the latter are assumed to be completely general in nature, and not necessarily produced by transformations such as projectivities between images. See Besl and McKay [11] for some work on the problem of registering images based upon threedimensional shapes. Closer to the automated homology that we seek are the algorithms of Grenander and Miller [77]. The approach of Grenander and Miller is part of a larger program of interpretation and representation of complex images using the Pattern Theory pioneered by Ulf Grenander, and briefly surveyed in that paper. We shall examine the similarities and differences between the methods later. For shape analysis, suppose that we have a collection of images of different objects, say images of brooches or perhaps images of faces, that vary slightly but not excessively in shape. Let us assume that in all images we are looking at essentially the same type of object, so that between any two images an approximate homology can in principle be established. For the purposes of analysis, we perform a rough standardization on the images so that all images have the same dimensions in pixel units, and so that each object within the image is centered and standardized in terms of orientation and scale. This last requirement is not required to be accomplished with careful Procrustean matching. Rather we will assume that homologous points between images are separated by small distances compared to the dimensions of the images. To simplify further, we also suppose that the images are grayscale or dithered black and white pictures such as can be produced by many image viewers.
6.5.2
Automated Block Homology
To describe automated homology, let us consider two images. Suppose that we wish to establish a homology from one to the other. This should be a function defined on all the pixel locations of one image and mapping to the pixel locations of the other. However, in practice we could choose a smaller set of locations by superimposing a rectangular lattice of points over each image. Equivalently, we can partition the images into blocks and suppose that these lattice points are the centers of the blocks. Our task is then to construct a correspondence between the lattice points that most nearly corresponds to the homology between the images. Consider a function h = (h lrh2 ) that maps a point at the jth row andkth column
6.5
\
Automated Homology
• • •
195
• • •
4
/
• •
./
•
/
• •
•
FIGURE 6.7. Block assignment for automated homology. In order to construct an automated homology between images, the images are transformed to matrices and then divided into blocks. A first step in the construction of the homology is to find a mapping between the blocks in the first image and those in the second so that a measure of discrepancy W is minimized.
of one image to the h1(j, k)th row and h2(j, k)th column of the other as in Figure 6.7. Let W[j, k; l, m] be a measure of mismatch of the homology between location (j, k) of the first image and location (l, m) of the second image. To construct the best homology from the lattice on one image to the lattice of the other we can minimize
LLW[j,k; h1(j,k),h2(j,k)] j
(6.25)
k
over all functions h. We do not require that h be a 11 function, as this is much too restrictive for our purpose. The next step that needs to be considered is the construction of the mismatch function W[j, k; l, m]. To construct such a function, the images must be placed into an environment in which they can be quantitatively compared. A variety of packages are available at the time of writing to assist in the analysis. The description that follows represents an approach found useful in the analysis of the Iron Age brooches of Figures 1.1 and 3.7. The images are first transformed to matrices of real numbers byconversion to ASCII format. For a grayscale image, the entries in the matrix will denote the degree of darkness at a particular pixel location. For a dithered image, the matrix will consist of entries of zeros and ones corresponding to black and white pixel values. A standard tool for conversion of an image to a matrix is the XV viewer available on Xwindows terminals and the UNIX operating system. The matrices can then read into MATLAB, which provides special tools for the manipulation of matrices. Each matrix representing an image can then be subdivided into blocks of size p x p, say. We can think of the lattice points (j, k) as being centered in the middle of these blocks so that the mismatch W[j, k; l, m] between lattice point (j, k) in the first image and (l, m) in the second is interpreted as
196
6.
Some Examples of Shape Analysis
6.5
a measure of mismatch between their corresponding blocks. Thus we shall seek a function that measures the mismatch between block (j, k) in the first matrix and block (l, m) in the second. Suppose the matrix of the first image is divided up into blocks (A j k) where each A j k is itself a p x p matrix. Similarly, let us suppose that the matrix from the second image is divided up in blocks (Elm) that are also of the same dimension. Now, a function h mapping block A j k to block Elm is a candidate for a homology between the images. To measure the discrepancy between the images, we can find some measure of distance between the matrices A j k and Elm. If the entries of these matrices are zeros and ones a suitable measure of mismatch could be obtained by counting the number IIAjk  Elmll of discordant entries between them. More generally, since A j k and Elm are each p x p matrices, we can regard them as vectors in RPxP. The distance IIA j k  Blmll can then be taken as Euclidean distance between these vectors. An optimization algorithm can then be run using this choice of W. However, it should be noted that such an automated homology algorithm has no respect for the natural topology of the image. Points that are the centers of neighboring blocks in the first image could be mapped by the optimal h to opposite sides of the second image. To counter this trend we can introduce another term to the formula for W[j, k; l, m] that measures the distance between the coordinates. Thus our formula for W[j, k; l, m] becomes
(6.26) for appropriate weights Al ::::: 0 and A2::::: 0, and suitable nondecreasing functions 'Tj: R + + R +. The functions 'h and T2 can be chosen to be the identity functions. However, in many cases it seems reasonable to rule out transformations h that will effect too radical a distortion of the image. This can be accomplished by making 'h and T2 increase faster than linearly. To restrict the search to transformations that perturb points to a maximum distance of f > 0, say, a reasonable choice for T2 would be one for which ~12(x) = 00 when x > f. If A2 is large compared to Al then the optimal homology found will be inelastic because transformations close to the identity transformation will be favored. On the other hand, when Al is large compared to A2 the algorithm will allow more drastic transformations to match features in the images. Having established a preliminary correspondence between blocks in two images, we still have to construct a full homology between the images by smoothly extending from the correspondence between the centers of the blocks to the rest of the image. We begin by assigning a vector Vjk to the center Xjk of each block A j k pointing from Xjk to the center of the corresponding block Eh(j,k) in the second image. Thus Vjk can be regarded as a vector field on the lattice of centers Xjk of blocks. The vector field is smoothed across the entire image by assigning a vector to
Automated Homology
197
each point x in the first image whose value is a weighted average of neighboring lattice vectors Vjk. This smoothing can be performed using a Gaussian kernel. Many other choices are also reasonable. Associated with the vector field v(x) on the first image is a transformation h' from the first image to the second that has this vector field as its field of difference vectors. So, we can write h'(x) = x + v(x). We could stop at this point and let this transformation h' be the required homology. However, this construction, while transforming the shape of the first image closer towards the second, still tends to be too rough an approximation to the desired homology to be suitable for shape analysis. This would appear to be due to the coarseness of the block size p x p, which has to be sufficiently large to allow the discrimination of salient features in the images. To compensate for this, we can allow the features of the first image to undergo a small displacement in the direction of the initial homology h'. This is achieved by shrinking the displacement vector so that a point x is mapped to the new point x + cv(x) for f > O. The choice of e = 1 represents the full transformation by h'. However, this transformation is too drastic in some cases. For these cases, a more modest perturbation is preferred with some E < 1. The perturbed image then replaces the original first Jmage and is partitioned into blocks. A homology hI/ is then constructed in a similar fashion between the perturbed first image and the second image. This procedure continues with hili, hI/II, '" until a satisfactory transformation (the composition of the small perturbations) from the original first image to the second image has been achieved.
6.5.3
An Application to Three Brooches
Michael Lewis has developed and implemented these techniques for an automated block homology routine. In Figure 6.8 we see an automated homology between the three brooches of Figure 1.1. There are a number of weighting factors and choices to be made by the researcher, such as the choice of Al and A2 above. A certain amount of standardization of the images must be done in advance. This need not be very precise, and can be incorporated into the automatic search procedure. However, it is advisable to have the researcher interacting with the procedure at this level as well, as the fitting is accomplished with any computer interpretation of the images as a whole. Thus the algorithm is a compromise between the expert selection and spline interpolation methods of Bookstein and a fully automated procedure that would be expected in computer vision. The latter is typically context sensitive. Perhaps a more precise description of the method would be to call it computer assisted homology.
198
6.
6.6 Notes
Some Examples of Shape Analysis
6.6
.199
Notes
This chapter has considered a few examples of the application of shape analysis. While the theme of shape is common to all examples, the methodologies used are quite diverse. This is even more the case for the entire range of applications in the literature. For this reason, a single chapter of applications cannot do justice to the variety of techniques and examples. The reader who is interested in particular applications will find, grouped by topic below, some references that can serve as a starting point for the exploration of the literature. I have taken the liberty of including examples in which shape plays an important role without necessarily being the primary topic of concern. Such applications leave room for future work involving the theory of shape.
6.6.1
Anthropology, Archeology, and Paleontology
II], [2], [7], [9], [10], [25J,[26]' [30], [31], [32J, [33], [44J, [45J, [58J, [59], [95], [106J, [107J, [109J, [110], [129], [130], [131], [133J, [155], [165], [175J, [177J.
6.6.2
Biology and Medical Sciences
For a more complete list of biomedical applications, the reader is referred to the references given in [24J.
FIGURE 6.8. Automated homology for three Iron Age brooches. The images of the three brooches can be seen along the diagonal of the 3 x 3 matrix of diagrams. Off the main diagonal, in the (j, k)th position, is a quiver diagram for the vector field of displacements for the homology that attempts to transform the jth brooch into the kth brooch. The vectors displayed are not the displacement vectors described above, but father rescaled versions of these, shrunk for convenience of graphical presentation. The 3 x 3 matrix of images has a natural antisymmetry properly: the sum of the vector fields in the (j, k) th and (k, j) th positions is zero. Blocks of size p x p = 16 x 16 were used in the partition of the images.
[10], [13], [14J, [15], [16], [17]' [18], [19], [20], [21], [24), [25), [26], [27], [45), [48J, [49], [50], [62], [63], [68J, [69J, [85], [106], [107J, [108],[126], [131J, [142], [143], [145], [152]' [169], [170], [172], [176J, [181], [182J.
6.6.3
Earth and Space Sciences
[28]' [29] ,[37], [42], [110], [118], [130], [173], [185].
6.6.4
Geometric Probability and Stochastic Geometry
[3], [4], [5], [6], [29], [30], [33], [46], [47], [89J, [91], [93], [95], [96J, [97], [100], [101], [102], [103], [111]' [112]' [118], [119], [128], [135], {141], {147]' [148], [153], [154], [155], [157], [159], [160J, [161], [166], [167], [171], [178], [179], [183].
6.6.5
Industrial Statistics
[11]' [40], [54], [84].
200
6.6.6
6. Some Examples of Shape Analysis
Mathematical Statistics and Multivariate Analysis
[S], [35], [3S], [52], [53], [55], [56], [57], [60], [66], [67], [70], [71], [72], [74], [75], [SO], [81], [86], [88], [89], [90], [91], [92], [93], [94], [95], [98], [99], [101], [102], [103], [104], [105], [Ill], [112], [113], [114], [116], [117], [123], [124], [134], [139], [140], [144], [150], [151], [153], [154], [159], [166], [168].
Bibliography
6.6. 'l Pattern Recognition, Computer Vision, and Image Processing [11], [36], [41], [77], [83], [148].
6.6.8
Stereo logy and Microscopy
[47], [48], [49], [50], [62], [147], [160], [161], [180], [181], [182].
6.6.9
Topics on Groups and Invariance
[12], [36], [120], [127], [132], [136], [137], [138], [156], [162], [184]. [1] Alexander, R. M. "Estimates of speeds of dinosaurs." Nature 261 (1976), 129130. [2] Alexander, R. M. "How dinosaurs ran." Scientific American 264 (1991),
130136. [3] Alexandrian, A. R. "Random quadrangular shapes by factorization." In Stochastic Geometry, Geometric Statistics, Stereology, TeubnerTexte zur Mathematik 65, ed. R. V. Ambartzumian and W. Weil. Teubner, Leipzig
(1984), 713. [4] Ambartzumian, R. V. "Random shapes by factorization." In Statistics in Theory and Practice, ed. B. Ranneby. Umea: Section of Forest Biometry, Swedish University of Agricultural Sciences (1982), 3541. [5] Ambartzumian, R. V. Factorization Calculus and Geometric Probability. Cambridge University Press, Cambridge (1990). [6] Baddeley, A. "Stochastic geometry: an introduction and reading list." Int. Statist. Rev. 50 (1982),179193. [7] Barker, P. A. "An exact method for describing metal weapon points." In Computer Applications in Archaeology, Proc. Annual Conf. Computer
202
Bibliography
Centre, University of Birmingham (1975), 38. [8] BarndorffNielsen, O. E., Blaesild, P., and Eriksen, P. S. Decomposition and invariance of measures, and statistical transformation models. Springer Lecture Notes in Statistics 58, SpringerVerlag, New York (1989). [9] Barrett, J., Bradley, R., Bowden, M., and Mead, B. "South Lodge after Pitt Rivers." Antiquity 57 (1983), 193204. [10] Benson, R. H., Chapman, R. E., and Siegel, A. F. "On the measurement of morphology and its change." Paleobiology 8 (1982), 328339. [11] Besl, P. J. and McKay, N. D. "A method for registration of 3d shapes." IEEE Transactions on Pattern Analysis and Machine Intelligence 14 (1992), 239256. [12] Bondesson, L. "Characterizations of probability laws through constant regression." Z. Wahrscheinlichkeits. verw. Gebiete 30 (1974), 93115. [13] Bookstein, F. L. The measurement of biological shape and shape change. Lecture Notes in Biomathematics 24. SpringerVerlag, New York (1978). [14] Bookstein, F. L. "Linear machinery for morphological distortion." Comput. Biomed. Res. 11 (1978), 435458. [15] Bookstein, F. L. "Foundations of morphometries." Ann. Rev. Ecol. Syst. 13 (1982), 451470. [16] Bookstein, F. 1. "A statistical method for biological shape comparisons." J. Theor. Biol. 107 (1984), 475520.
Bibliography
203
Information Processing and Medical Imaging, ed. S. 1. Bacharach. Nijhoff, Dordrecht (1986), 118. [21] Bookstein, F. 1. "Soft modeling and the measurement of biological shape." In Theoretical Empiricism: A General Rationale for Scientific ModelBuilding, ed. H. Wold. Paragon Press, New York (1987). [22] Bookstein, F. L. "Discussion of 'A survey of the statistical theory of shape' by D. G. Kendall." Statist. Sci. 4 (1989),99105. [23] Bookstein, F. L. "Principal warps: thinplate splines and the decomposition of deformations." IEEE Transactions on Pattern Analysis and Machine Intelligence 11 (1989), 567585. [24] Bookstein, F. L. Morphometric Tools For Landmark Data: Geometry and Biology. Cambridge University Press, Cambridge (1991). [25] Bookstein, F. L., Chernoff, B., Elder, R., Humphries, J., Smith, G., and Strauss, R. Morphometries in Evolutionary Biology: The Geometry of Size and Shape Change with Examples from Fishes. Academy of Natural Sciences of Philadelphia, Philadelphia (1985). [26] Bookstein, F. L. and Reyment, R. A. "Microevolution in Miocene Brizalina (Foraminifera) studied by canonical variate analysis and analysis of landmarks." Bull. Math. Biol. 51 (1989), 657679. [27] Bookstein, F. L. and Sampson, P. D. "Statistical models for geometric components of shape change." Comm. Statist. Theory Meth. 19 (1990), 19391972.
[17] Bookstein, F. L. "Tensor biometrics for changes in cranial shape." Ann. Human BioI. 11 (1984), 413437.
[28] Bookstein, F. L., Sampson, P. D., Lewis, S., Guttorp, P., and Hurley, C. B. "Computation and interpretation of deformations for landmark data in morphometries and environmetrics." Computing Sc. and Statist. Proc. of the 23th Symposium on the Interface (1991), 534541.
[18] Bookstein, F. L. "Transformations of quadrilaterals, tensor fields, and morphogenesis." In Mathematical Essays on Growth and the Emergence of Form, University of Alberta Press, Edmonton (1985), 221265.
[29] Boots, B. N. "Delaunay triangles: an alternative approach to point pattern analysis." Proc. Assoc. Amer. Geographers 6 (1974), 2629.
[19] Bookstein, F. L. "Size and shape spaces for landmark data in two dimensions (with discussion)." Statist. Sci. 1 (1986), 181242.
[30] Bradley, R. and Fulford, M. "Sherd size in the analysis of occupation debris." Bull. Univ. London Inst. Archaeology 17, 8594.
[20] Bookstein, F. L. "From medical images to the biometrics of form." In
[31] Bradley, R. and Small, C. "Looking for circular structures in post hole
204
Bibliography
Bibliography
205
distributions: quantitative analysis of two settlements from Bronze Age England." J. Archaeol. Sci. 12 (1985), 285297.
[44] Cogbill, S. "Computer post hole analysis with reference to the Bronze Age." Computer Appl. Archaeol. 1980, University of Birmingham (1980), 3538.
[32] Bradley, R. and Small, C. "Statistical analysis of structures at two settlements from Bronze Age England." MASCA J.4 (1986), Museum Applied Science Center for Archaeology, University of Pennsylvania, Philadelphia, 8695.
[45J Corner, B. D. and Richtsmeier, J. T. "Morphometric analysis of craniofacial growth in Cebus apella." Amer. J. Phys. Anthrop. 84 (1991), 323342.
[33] Broadbent, S. "Simulating the ley hunter (with discussion)." J. Roy. Statist. Soc. Ser. A 143 (1980), 109140. [34] Bronshtein, I. N. and Semendyayev, K. A. Handbook of Mathematics. Van Nostrand Reinhold, New York (1985). [35] Brown, G. W. "Reduction of a certain class of composite statistical hypotheses." Ann. Math. Statist. 11 (1940), 254270. [36] Bumcrot, R. J. Modern Projective Geometry. Holt, Rinehart and Winston, New York (1969). [37] Cahn, J. W. "The generation and characterization of shape." Adv. Appl. Probab. Special Supplement (1972), 221242. [38] Carne, T. K. "The geometry of shape spaces." Proc. London Math. Soc. 61 (1990), 407432. [39] Carroll, L. Curiosa Mathematica, Part II: Pillow Problems. Macmillan, London (1893). [40] Chen, G. and Chen, J. "Geometric quality inspection." Submitted to Statistica Sinica (1996). [41J Chow, Y., Grenander, U., and Keenan, D. M. "HANDS, a pattern theoretic study of biological shapes." Technical Report, Division of Applied Math. Brown University, Providence (1988).
[46] Cover, T. and Efron, B. "Geometrical probability and random points in a hypersphere." Ann. Math. Statist. 38 (1967), 213220. [47] Cowan, R. "A collection of problems in random geometry." In Stochastic Geometry, Geometric Statistics, Stereology, TeubnerTexte zur Mathematik 65, Edited by R. V. Ambarzumian and W. Wei!. Leipzig (1984), 6468. [48] CruzOrive, L.M. "Particle sizeshape distributions: the general spheroid problem, part 1." J. Microsc. 107 (1976), 235253. [49] CruzOrive, L.M. "Particle sizeshape distributions: the general spheroid problem, part 2." J. Microsc. 112 (1978), 153167. [50] CruzOrive, L.M. and Weibel, E. R. "Recent stereological methods for cell biology: a brief survey." Am. J. Physiol. 258 (1990) LI48LI56. [51] Dieudonne J. Treatise on Analysis, Vol. 3. Academic Press (1969). [52] Dryden, I. L. "The statistical analysis of shape data." PhD Thesis, Dept. of Statistics, University of Leeds, Leeds (1989). [53] Dryden, I. L. and Mardia, K. V. "General shape distributions in the plane." Adv. Appl. Probab. 23 (1991), 259276. [541 DuPuis, P. and Oliensis, J. "An optimal control formulation and related numerical methods for a problem in shape reconstruction." Ann. Appl. Probab. 4 (1994), 287346.
[42J Christaller, W. Die zentralen Orie in Silddeutschland Jena (1933). Translated by C. C. Baskin, Central Places in Southern Germany, PrenticeHall, New Jersey (1966).
[55J Fisher, N. I. Statistical Analysis of Circular Data. Cambridge University, Cambridge (1993).
[43] Chung, K. L. A Course In Probability Theory. Academic Press, New York (1974).
[56] Fisher, N. 1., Lewis, T., and Embleton, B. J. J. Statistical Analysis of Spherical Data. Cambridge University, Cambridge (1987).
206
Bibliography
[57] Fisher, R. A. "Two new properties of mathematical likelihood," Proc. Roy; Soc. London Ser. A 144 (1934), 285307. [58] Fletcher, M. and Lock, G. "Computerised pattern perception with post hole distributions." Sci. Archaeol. 23 (1981), 1520. [59] Fletcher, M. and Lock, G. "Post built structures at Danebury hillfort: an analytical search method with discussion." Oxford J. Archaeol. 6 (1984), . 175196. [60] Flury, B. A. "Principal points." Biometrika 77 (1990), 3341. [61] Fllkugawa, H. and Pedoe, D. Japanese Temple Geometry Problems. Charles Babbage Research Centre, Winnipeg (1989). [62] Girling, A. J. "Shape analysis for the anisotropic corpuscle problem." J. Roy. Statist. Soc. Ser. B 55 (1993), 675686. [63] Goodall, C. R. "The statistical analysis of growth in two dimensions." PhD Dissertation, Department of Statistics, Harvard University (1983). [64] Goodall, C. R. "The growth of a twodimensional figure: strain crosses and confidence regions." Proc. Statist. Comput. Sect. Amer. Statist. Assoc.
Bibliography
207
density." Adv. Appl. Probab. 23 (1991), 496514. [71] Goodall, C. R. and Mardia, K. V. "The noncentral Bartlett decompositions and shape densities." J. Multivariate Anal. 40 (1992),94108.
[72] Goodall, C. R. and Mardia, K. V. "Multivariate aspects of shape theory." Ann. Statist. 21 (1993), 848866. [73] Gough, J. and Mardia, K. V. "Shape modelling using deformable polygons." Technical Report. Department of Statistics, University ofLeeds (1990). [74] Gower, J. C. "Some distance properties oflatent root and vector methods used in multivariate analysis." Biometrika 53 (1966), 325338. [75] Gower, J. C. "Generalized Procrustes analysis." Psychometrika 40 (1975), 3350. [76] Gregory, R. L. and Gombrich, E. H. Illusion. in Art and Nature. Duckworth, London (1973).
(1984), 165169.
[77] Grenander, U. and Miller, M. 1. "Representation of knowledge in complex systems." J. Roy. Statist. Soc. Ser. B 56 (1994), 549603.
[65] Goodall, C. R. "Discussion of 'Size and shape spaces for landmark data in two dimensions' by F. L. Bookstein." Statist. Sci. 1 (1986), 181242.
[78] Guillemin, V. and Pollack, A. Differential Topology. Prentice Hall, New Jersey (1974).
[66] Goodall, C. R. "Procrustes methods in the statistical analysis of shape (with discussion)." J. Roy. Statist. Soc. Ser. B 53 (1991), 285339.
[79] Hewitt, E. and Ross, K. A. Abstract Harmonic Analysis, Vols. I, II. SpringerVerlag, Berlin (1963).
[67] Goodall, C. R. and Bose, A."Procrustes techniques for the analysis of shape and shape change." In Computer Science and Statistics, ed. R. M. Heiberger. American Statistical Association, Alexandria (1987), 8692.
[80] Heyer, H. Probability Measures on Locally Compact Groups. SpringerVerlag, Berlin (1977).
[68] Goodall, C. R. and Green, P. B. "Quantitative analysis of surface growth." Bot. Gaz. 147 (1986),115. [69J Goodall, C. R., Lange, N. and Moss, M. L. "Growthcurve models for
[81] Hills, M. "Allometry." In Encyclopedia of Statistical Sciences Vol. 1. Wiley, New York (1982), 4854. [82] Hogg, R. V. and Craig, A. T. Introduction to Mathematical Statistics. Fourth Edition. Macmillan, New York (1978).
repeated triangular shapes." Manuscript. [70] Goodall, C. R. and Mardia, K. V. "A geometric derivation of the shape
[83] Horn, B. K. P. and Brooks, M. J. Shape From Shading. MIT Press, Cambridge (1989).
208
Bibliography
. Bibliography
209
[84] Hulting, F. 1.. "Methods for the analysis of coordinate measurement data." Computing Science and Statistics 24 (1992), 160169.
Physical Systems, ed. D. Welsh and G. Grimmet. Oxford University Press, .Oxford (1990), 203217.
[85] Huxley, J. S. Problems of Relative Growth, 2nd ed. Dover, New York (1972).
[98] Kent, J. T. "The complex Bingham distribution and shape analysis." J. Roy. Statist. Soc. Ser. B 56 (1994), 285299.
[86] Jones, M. C. and Sibson, R. "What is projection pursuit? (with discussion)." J. Roy. Statist. Soc. Ser. A 150 (1987), 136.
[99] Kovalenko, 1. N. "On recovering the additive type of a distribution over a sequence of runs of independent observations." Trudy Vsesoyuzn. Soveshcheniya po Teorii Veroyatnostei i Matematicheskoi Statistike, Erevan (1960), 148~159. In Russian.
[87] Kendall, D. G. "The diffusion of shape." Adv. Appl. Probab. 9 (1977), 428430. [88] Kendall, D. G. "The statistics of shape." In Interpreting Multivariate Data, ed. V. Barnett. Wiley, New York (1981), 7580. [89] Kendall, D. G. "The shape of PoissonDelaunay triangles." In Studies ·in Probability and Related Topics, ed. M. C. Demetrescu and M. Iosefescu. Nagard, Sophia (1983). [90] Kendall, D. G. "Shape manifolds, Procrustean metrics, and complex projective spaces." Bull. London Math. Soc. 16 (1984), 81121. [91] Kendall, D. G. "Exact distributions for shapes of random triangles in convex sets." Adv. Appl. Probab. 17 (1985), 308329. [92] Kendall, D. G. "A survey of the statistical theory of shape." Statist. Sci. 4 (1989), 87120.
[100] Langford, E. "Probability that a random triangle is obtuse." Biometrika 56 (1969), 689690. [101] Le, H.L. "Explicit formulae for polygonally generated shapedensities in the basic tile." Math. Proc. Camb. Philos. Soc. 101 (1987), 313321.
[102] Le, H.L. "Singularities of convexpolygonally generated shapedensities." Math. Proc. Camb. Philos. Soc. 102 (1987), 587596. [103] Le, H.L. "A stochastic calculus approach to the shape distribution induced by a complex normal model." Math. Proc, Cambridge Philos. Soc. 109 (1990), 221228. [104] Le, H.L. "On geodesics in Euclidean shape spaces." J. London Math. Soc. 44 (1991), 360372.
[93] Kendall, D. G. "Random Delaunay simplexes in R'"." J. Statist. Plann. Inf. 25 (1990), 225234.
[105] Le, H.L. and Kendall, D. G. "The Riemannian structure of Euclidean spaces: a novel environment for statistics." Ann. Statist. 21 (1993), 12251271.
[94] Kendall, D. G. "Spherical triangles revisited." The Art of Statistical Science. A Tribute to G. S. Watson, ed. K. V. Mardia (1992). Wiley, New York, 105113.
[106] Lele, S. R. "Some comments on coordinate free and scale invariant methods in morphometries." Amer. J. Phys. Anthrop. 85 (1991), 407418.
[95] Kendall, D. G. and Kendall, W. S. "Alignments in twodimensional random sets of points." Adv. Appl. Ptobob. 12 (1980), 380424.
[107] Lele, S. R. and Richtsmeier, J. T. "Euclidean distance matrix analysis: a coordinatefree approach for comparing biological shapes using landmark data." Amer. J. Phys. Anthrop. 86 (1991), 415427.
[96] Kendall, W. S. "Symbolic computation and the diffusion of shapes of triads." Adv. Appl. Probab. 20 (1988), 775797. [97] Kendall, W. S. "The diffusion of Euclidean shape." In Disorder in
[108] Lele, S. R. and Richtsmeier, J. T. "On comparing biological shapes: detection of influential landmarks." Amer. J. Phys. Anihrop. 87 (1992), 4965.
210
Bibliography
Bibliography
[109] Litton, C. D. and Restorick, J. "Computer analysis of post hole distri
211
Bartlett, Boston (1992).
butions." 'Computer Applic. Archaeology 1983, University of Birmingham (1983), 8592. [110] Lohmann, G. P. "Eigenshape analysis of microfossils: a general morphometric procedure for describing changes in shape." Math. Geol. 15 (1983), 659672. [111] Mack, G. "Expected number of aggregates in a random distribution of n points." Proc. Cambridge Philos. Soc. 46 (1950), 285292. [112] Mannion, D. "A Markov chain of triangle shapes." Adv. Appl. Probab. (1988), 348370. [113] Mardia, K. V. Statistics of Directional Data. Academic Press, London (1972). [114] Mardia, K. V. "Shape analysis of triangles through directional techniques." J. Roy. Statist. Soc. Ser. B 51 (1989), 449~458. [115] Mardia, K. V. "Discussion of 'A survey of the statistical theory of shape' by D. G. Kendall." Statist. Sci. 4 (1989), 108J11. [116] Mardia, K. V. and Dryden, 1. L. "Shape distributions for landmark data." Adv. Appl. Probab. 21 (1989), 742755. [117] Mardia, K. V. and Dryden, 1. L. "The statistical analysis of shape data." Biometrika 76 (1989), 271281. [118] Mardia, K. V., Edwards, R., andPuri, M. L. "Analysis of central place theory." Bull. Int. Statist. Inst. 47 (1977), 93110. [119] Miles, R. "On the homogeneous Poisson process." Math. Biosciences 6 (1970), 85:127. [120] Montgomery, D. and Zippin, L. Topological Transformation Groups. Interscience Publishers, New York (1955). [121] Morgan, F. Geometric Measure Theory: A Beginner's Guide. Academic Press, Boston (1988). [122] Morgan, F. Riemannian Geometry: A Beginner's Guide. Jones and
[123] Mosier, C. 1. "Determining a simple structure when loadings for certain tests are known." Psychometrika 4 (1939), 149162. [124] Mosimann, J. E. "Size allometry: size and shape variables with characterizations of the lognormal and generalized gamma distributions." J. Amer. Statist. Assoc. 65 (1970), 930948. [125] Mosimann, J. E. "Size and shape analysis." In Encyclopedia of Statistical Sciences Vol. 8. Wiley, New York (1988), 497507. [126] Mosimann, J. E. and James, F. C. "New statistical methods for allometry with application to Florida redwinged blackbirds." Evolution 33 (1979), 444459. [127] Neyman, J. and Scott, E. "Consistent estimates based on partially consistent observations." Econometrics 16 (1948), 132. [128] Okabe, A., Boots, B., and Sugihara, K. Spatial Tessellations Concepts and Applications of Voronoi Diagrams. Wiley, New York (1992). [129] Orton, C. Mathematics in Archaeology. Cambridge University, Cambridge (1980). [130] Ostrom, J. H. "Were some dinosaurs gregarious?" Paleogeog. Paleoclimat. Paleoecol. 11 (1972), 287301. [13l] Oxnard, C. E. The Order of Man: A Biomathematical Anatomy of the Primates. Yale University, New Haven (1984). [132] Parthasarathy, K. R. Probability Measures on Metric Spaces. Academic Press, New York (1967). [133] Paul, G. S. Predatory Dinosaurs of the World: A Complete Illustrated Guide. Simon and Schuster, New York (1989). [134] Pitman, E. J. G. "The estimation of the location and scale parameters of a continuous population of any given form." Biometrika 30 (1939),391421. [135] Portnoy, S. "A Lewis Carroll pillow problem: probability of an obtuse
212
Bibliography
triangle." Statist. Sc. 9 (1994), 279284. [136] Prakasa Rao, B. L. S. "On a characterization of probability distributions on locally compact abelian groups." Z. Wahrscheinlichkeits. verw. Gebiete 9 (1968), 98100. [137] Prohorov, Yu. V. "On a characterization of a class of probability distributions by distribution of some statistics." Theor. Probab. Appl. 10 (1965), 438445.
Bibliography
213
[148] Serra, J. P. Image Analysis and Mathematical Morphology. Academic Press, New York (1982). [149] Schwerdtfeger, H. Geometry of Complex Numbers. Dover, New York (1962). [150] Sibson, R. "Studies in the robustness of multidimensional scaling: Procrustes statistics." J. Roy. Statist. Soc. Ser. B 40 (1978), 234238.
[138] Rao, C. R. Linear Statistical Inference and Its Applications. Wiley, New York (1965).
[151] Sibson, R. "Studies in the robustness of multidimensional scaling: perturbational analysis of classical scaling." J. Roy. Statist. Soc. Ser. B 41 (1979),217229.
[139] Reyment, R. A. Blacklith, R. E., and Campbell, N. A. Multivariate Morphometries, 2nd ed. Academic Press, New York (1984).
[152] Siegel, A. F. and Benson, R. H. "A robust comparison of biological shapes." Biometrics 38 (1982), 341350.
[140] Ripley, B. D. Spatial Statistics. Wiley, New York (1981).
[153] Silverman, B. W. and Brown, T. C. "Short distances, flat triangles and Poisson limits." J. Appl. Probab. 15 (1978), 815825.
[141] Ripley, B. D. arid Rasson, J.P. "Finding the edge of a Poisson forest." J. Appl. Probab. 14 (1977), 483491. [142] Rohlf, F. J. and Bookstein, F. L. Proceedings of the Michigan Morphometries Workshop. The University of Michigan Museum of Zoology, Ann Arbor (1990). Special Publication Number 2. [143] Rohlf, F. J. and Slice, D. "Methods for comparison of sets of landmarks." Syst. Zool. 39 (1990), 4059.
[154] Small, C. G. "Distributions of shape and maximal invariant statistics." PhD Thesis. University of Cambridge, Cambridge (1981). [155] Small, C. G. "Random uniform triangles and the alignment problem." Math. Proc. Camb. Philos. Soc. 91 (1982),315322. [156] Small, C. G. "Characterization of distributions from maximal invariant statistics." Z. Wahrsch. verw. Gebiete 63 (1983), 517527.
[144] Rukhin, A. "Charakterisierung del' Transformationsparameterfamilie." Z. Wahrscheinlichkeits. verw. Gebiete 38 (1977), 287291.
[157] Small, C. G. "A classification theorem for planar distributions based on the shape statistics of independent tetrads." Math. Proc. Camb. Philos. Soc. 96 (1984), 543547.
[145] Sampson, P. D. "Dental arch shape: a statistical analysis using conic sections." Amer. J. Orthodont. 79 (1981), 535548.
[158] Small, C. G. "Discussion of' A survey of the statistical theory of shape' by D. G. Kendall." Statist. Sci. 4 (1989), 105108.
[146J Sampson, P. D. and Siegel, A. F. "The measure of 'size' independent of 'shape' for multivariate lognormal populations." J. Amer. Statist: Assoc. 80 (1985), 910914.
[159J Small, C. G. "Techniques of shape analysis on sets of points." Int. Statist. Rev. 56 (1988), 243257.
[147J Santalo, L. A. Integral Geometry and Geometric Probability. Encyclopedia of Mathematics and Its Applications. AddisonWesley, Reading (1976).
[160] Small, C. G. "Reconstructing convex bodies from random projected images." Can. J. Statist. 19 (1991), 341347. [161] Small, C. G. "A counterexample to a conjecture on random shadows."
214
Bibliography
Can.J. Statist. 20 (1992), 463468. [162] Small, C. G. and Murdoch, D. J. "Nonparametric NeymanScott problems: telescoping product methods." Biometrika 80 (1993), 763779. [163] Spivak, M. A Comprehensive Introduction to Differential Geometry, Volume 1. Publish or Perish, Berkeley (1970). [164] Sprent, P. "The mathematics of size and shape." Biometrics 28 (1972), 2337. [165] Stone, J. F. S. "The DeverelRimbury settlement on Thorny Down, Winterbourne gunner, S. Wilts." Proc. Prehistoric Soc. 7 (1941), 114133.
Bibliography
215
[175] Wainwright, R. A Guide to the Prehistoric Remains in Britain, Volume I: South and East. Constable, London (1978). [176] Wainwright, S. A. Axis and Circumference: The Cylindrical Shape of Plants and Animals. Harvard University Press, Harvard (1988). [177] Watkins, A. The Old Straight Track. Methuen (1925). Republished Garnstone Press (1970). [178] Watson, G. S. "Random triangles." In Proc. Conf. Stochastic Geometry, ed. O. BarndorffNielsen. Aarhus University, Aarhus (1983). [179] Watson, G. S. "The shape of a random sequence of triangles." Adv. Appl. Probab, 18 (1986), 156169.
[166] Stoyan, D. "Estimation of distances and variances in.Bookstein'sIandmark model." Biometr. J. 32 (1990),843849.
[180] Weibel, E. R. Stereological Methods. Academic Press, London (1980).
[167] Stoyan, D., Kendall, W. S., and Mecke, J. Stochastic Geometry and Its Applications. Akademie, Berlin (1987).
[181] Wicksell, S. D. "The corpuscle problem, part 1." Biometrika 17 (1925), 8499.
[168] Stoyan, D. and Stoyan, H. "A further application of D. G. Kendall's procrustes analysis." Biometr. J. 32 (1990), 293301.
[182] Wicksell, S. D. "The corpuscle problem, part 2." Biometrika 18 (1926), 152172.
[169] Strauss, R. E. and Bookstein, F. L. "The truss: body form reconstruction in morphometries." Systematic Zool. 31 (1982), 113135.
[183] Ziezold, H. "On expected figures in the plane." Math. Res. Ser. 51 (1989), 105110.
[170] Stuetzle, W., Gasser, T. Molinari, L., Largo, R. H., Prader, A., and Huber, P. J. "Shape invariant modelling of human growth." Ann. Human Biol. 7 (1980), 507528.
[184] Zinger, A. A. and Linnik, Yu. V. "On characterizations of the normal distributions." Theor. Probab. Appl. 9 (1964), 624626.
[171] Sukiasian, H. S. "Two results on triangle shapes." In Stochastic Geometry, Geometric Statistics, Stereology, TeubnerTexte zur Mathematik 65, ed. R. V. Ambartzumian and W. Weil. Teubner, Leipzig (1984), 21O~221.
[172] Thompson, D'A. W. On Growth and Form. Cambridge University Cambridge (1961). Reprinted from 1917. [173] Tobler, W. R. "The comparison of plane forms." Geograph. Anal. 10 (1978), 154162. [174] Wagner, H. M. Principles of Operations Research. PrenticeHall, Englewood Cliffs (1975).
[185] Zuiderwijk, E. J. "Alignment of randomly distributed objects." Nature 295 (1982), 577578.
Index
Acceptance function, 184 Acceptance method for simulation, 170 Acute triangle, 150 Affine connection, 50 Affine transformation, see Transformations Aligned landmarks, 58, 156, 161 see also Collinear points or landmarks Alignments of megalithic stones, 4, 158, 160 Allometry, 2, 4, 26, 113 growth allometry, 5, 112 Amphorae, 3 Angular criterion for alignment, 160 Anisotropy, 95 local anisotropy, 110112 loganisotropy, 9596, 98, 105 Annulus, 23, 187190 Anthropology, 199 Antipodal points on a sphere, 5556, 58, 7371, 76, 128129 characterizing geodesics using antipodal points, 77
Arc length in Riemannian manifold, 4850, 52 Archeology, 12, 23, 94, 158, 199 ASCII format, 195 Astronomy, 2 Automated homology, see Homology Axial data, 146 Axis, used to define orientation, 10
Bending energy, 108 Bertrand's paradox, 4 Bilateral symmetry, 181 Binomial distribution, see Distribution Binomial process, see Point process Biology, 2, 2627, 95, 199 Blaschke constants, 163 Bookstein coordinates, see Shape coordinates Boolean combination of events, 118 Borel sets of a manifold, 119i21, 128 Boundary effect, 185186, 189 Boundary of a manifold, see Manifold
218
Index
Index
Boundary of a set, 145 Broadbent factor, 156, 163 see also Aligned landmarks Brooches from Iron AgeMiinsingen, 67, 9, 13, 24, 9294 application of principal coordinate analysis to brooch data, 9294 size versus shape analysis, 94 Brownian motion, 1
Calculus of variations, 49 Cartesian coordinates, 24, 51, 113 Cartesian product of manifolds, see Manifold, Cartesian product Casson spheres, see Shape manifold CauchyRiemann equations, 111 Centroid, 89, 77, 84, 180 Change of variables, see Transformations of statistics Characteristic equation for eigenvalues, see Matrix, eigenvalues Circle at infinity, 64 Circlepreserving property of Moebius transformation, 72 of stereographic projection, 72 Circularly symmetric density, 154, 156157 Circumcircle, 142, 145, 189 Circumradius, 168 Circumsphere, 141, 143 Closed complex plane, see Complex plane Closed set, 29, 37, 119 Colatitude on a 2sphere, 54, 123 Collinear points or landmarks, 156157 collinear triangles, 7475, 77, 97, 179 singularity sets and collinearities, 85 see also Alignments Commutative diagram, 128 Compact set, 3738 compactness of Kendall's shape spaces, 80
Complex analytic function, see Function Complex dimensions versus real dimensions, 59, 77 Complex lines through the origin, 5960, 7778 Complex plane, 12, 31, 69, 71, 77, 80,97 closed complex plane, 7174 point at infinity in complex plane, 71 Computer vision, 193194, 200 Configuration of particles, 185186 expected number of configurations, 185186 size of configuration, 185 Configuration of sample, 3 Confluent hypergeometric function, 167 Conformal transformation, see Transformation Congruence and congruent sets, 4, 27,35 Content of a set, see Volume Convex hull, 2930 Convex set, 29, 143, 163 Convexity, 27 Coplanar configurations of landmarks in R 3 , 82 Countable intersection of open sets, 119 Covariance matrix, 130 Craniumtojaw ratio of skulls, 16, 26 Curvature of a surface or manifold, 38 Curvilinear coordinates, 24, 26, 106,113
Delaunay simplex, 141, 143145, 168 presizeandshape distribution, 143145 shape distribution, see Distribution, shape distribution Delaunay tessellation, 141143, 146, 168
applied to central place theory, 146 applied to crystallography, 146 duality with Voronoi tessellation, 146 Delaunay triangle, 141142 Density function, see Distribution Differential geometry, vvi, 16, 36, 47, 59, 66, 78 Differential manifold, see Manifold Differential singularity, 85 Dimension reduction techniques, 88,91 Directed line, 135137, 148 Directional cosine, 51 Directional data, 146, 175 Directional median, 175176 see also Mt. Tom dinosaur tracks Distance matrix, see Matrix Distribution absolutely continuous distribution, see continuous distribution binomial distribution, 135, 143, 148 continuous distribution, 120, 123 density function, 123127, 132, 147 discrete distribution, 120 distribution function, 120, 147 induced probability distribution, 119121 invariant, 125129 marginal density, 125 normal on Euclidean space, 4, 6, 26, 79, 130131, 133 elliptical normal, 154156, 160161 spherical normal, 130131, 149151 normal on spheres Brownian motion distribution, 180 Fisher, 180 offset normal, see projected normal projected normal, 131134, 148149, 180 Poisson distribution, 138, 148
219
shape distributions concentration parameter for, 165 IID elliptical normal landmarks in R 2 , 155156 IID spherical normal planar landmarks, 149151 IID spherical normal planar landmarks in Bookstein coordinates, 151152 MardiaDryden density, 134, 152, 163167, 180 Miles' triangle density, 170,
172 PoissonDelaunay, 167170 uniform, 4, 23, 27, 123, 125, 133134, 137, 148149, 155, 162163, 188
eblunt triangle, 4, 160161, 171 Earth science, 199 Eigenvalues and eigenvectors, see Matrix ' Einstein summation convention, 47 Ellipse, 32, 171172 anisotropy of an ellipse, see Anisotropy image of circle under affine transformation, 9596 semirnajor axis, 32, 9596 semiminor axis, 32, 9596 stretch factor, see Stretch factor Ellipsoid, 32 principal axes, 32 Embedding, 38 Equilateral triangle, 74, 76, 155 Equivalence class complex projective space as set of e. classes, 5960, 77 shapes as e. classes of preshapes, 1112, 60, 77, 7980 tangent vectors as e. classes, 4246, 53, 67 Euclidean space, vi, 9, 16, 29, 3839, 41, 43, 62, 88, 92, 112, 130, 139 EulerLagrange equations, 49, 51 Event, 117, 134
220
Index
Expected value, 2324, 121 Exploratory analysis of shapes, 88 Exponential growth, 5
Factorization calculus, 186 Fractal,2 Frobenius norm, see Norm on the space of upper triangular matrices F'roude numbers, 1'16 FubiniStudy metric, see Metric, FubiniStudy Function complex analytic, 111 continuous, 3637, 120 covering from sphere to real projective space, 5658, 127129 derivative of, 36, 5253, 57, 86 differentiable or smooth, 36, 41, 5253, 57, 81, 83, 124 Hopf fibration, 7879, 83 projection, 83, 131132 as example of Riemannian submersion, 78 onto subspace spanned by eigenvectors, 91 Riemannian submersion as local orthogonal p., 86 submersion, 81, 8384, 8687 Riemannian submersion, 78, 84, 8687
Gamma function, 144 Gaussian curvature, see Manifold General position of landmarks, 99, 140 Geodesic distance, 12, 14, 48, 50, 5455, 5758, 60, 6364, 72, 78, 88, 91, 104105, 114, 116, 125, 128, 134, 179 Geodesic path, see Path in a manifold Geometric measure, 121124, 147, 166, 185 factorization of g. m., 185186 Geometr!c probability, 34, 199
Index Gradient, 38, 42, 112 GramSchmidt orthogonalization, 99 Grayscale image, 195 Great circle distance, 12, 54, 76 Great circle of collinear triangle shapes, see Shape manifold, sphere of triangle shapes Great circle of isosceles triangle shapes, see Shape manifold, sphere of triangle shapes Great circle on a sphere, see Path in a manifold Group of transformations, vi, 3335, 80, 128129, 200 center, 33, 129 commutative or Abelian, 3334 compact, 145 composition of transformations within g., 33 examples, see Transformations free action of g. and singularities, 8385 homotopy g., 11 identity transformation, 34, 83 inverse transformation, 33 isometry g., 125129 subgroup, 3334, 80 transitive, 126, 129, 147 trivial,33
Hairy ball theorem, 66 Half circle, 64 HeineBorel theorem, 37 Hermitian inner product, 32, 61 Heterogeneous scale changes, 112 High exponents, 187 Homogeneous function, 4 Homology, 24, 26, 95, 110 automated homology, 94, 193198 application to Iron Age brooches, 193, 197198 automated block homology, 194197 GrenanderMiller method, 194 mismatch function, 195
versus correspondence problem, 193194 versus Procrustean matching, 194 biological versus nonbiological, 24 eyes, as examples of, 24 homologous landmarks, 24, 107, 193 problem of homology, 2426, 35 registration of images, 194 relation to method of coordinates, 24 Hopf fibration, see Function Horizon at infinity, 64, 123 Horizontal geodesic, see Path in a manifold Horizontal tangent space, see Tangent vector, tangent space Hyperbolic half space, see Manifold Hyperplane, 34
Identically distributed statistics, 121 Image processing, 200 Imaginary part of a complex number, 12, 31, 6970, 72, 160 Independence, 121 Indicator random variable, 134 Induced probability distribution, see Distribution Industrial statistics, 199 Infinitesimal distance in UT(n), 102, 105 Inner product, 12, 32, 4748, 55, 61, 90, 167 Intensity of scattering, 190191 Interior of set, 29 Interpoint geodesic distance matrix, see Matrix, distance matrix Interpolation, 26, 107 Invariance, 200 invariance of landmarks under rotations, 84 invariance of metric tensor, see Metric tensor
221
invariance of uniform distribution, 125129 invariant function, 184 invariant measure, 137, 146 invariant statistic, 3 Iron Age brooches, see Brooches from Iron Age Isometries, see Transformations, isometries Isoperimetric inequality, 2 Isosceles triangle, 115
Jacobian, see Transformations Jacobian matrix, see Matrix
Kendall school of shape analysis, v, 2627
Labeled set or figure, 35 counterclockwise labeling of planar triangles, 97 Land's End, Old Stones of 4, 158163, 184 ley lines, 158, 184 scatterplot, 159 see also Alignments Landmarks, 2, 7, 9, 1114, 16, 2627, 58, 6970, 7677 Late Bronze Age people, 184 Length of an infinitesimal displacement, 43 Lens of eblunt triangles, 161 LeviCivita connection, 50 Ley lines, see Alignments of megalithic stones Linear fractional transformation, see Transformations, Moebius transformation Linearly independent vectors, 122 Local anisotropy, see Anisotropy Local isometry, 58, 113 Local shape variation, 111113 Location information, 711, 84, 99, 133 Location parameters, 3 Loganisotropy, see Anisotropy
Index
222 _ Index Logarithmic coord'inates, 5 Longitude on a 2sphere, 54, 123 Lung tissue, 2
Manifold, vi, 1, 38~40 atlas on m., 3841,51, 5455, 83 boundary of m., 67, 81 Cartesian product of manifolds, 5152, 55, 166 chart on m., 3841, 44, 51, 54, 56, 5960 complex coordinates on m., 59 complex projective space, 1, 12, 5962, 77~79, 88, 129 constant curvature, 114, 146 coordinates on m., 41, 4446, 5254, 60 curvature of m., 5~51, 114 cylinder as m., 135136, 148 differential m., 2, 37,39, 4143, 45, 5154, 56, 59, 118 extrinsic properties ofm., 41, 124 fiber bundle, 166 Gaussian curvature, 78 hyperbolic half space, 6266, 123 intrinsic properties of m., 41, 46 Klein bottle, 67 m.of negative curvature, 2, 62, 65 m.of positive curvature, 1, 62 m. with boundary, see boundary of a manifold Moebius strip, 67, 137, 148 patching criterion for charts, 3940 Poincare Disk, 6465, 147 Poincare Plane, 63, 65, 95, 99 Poincare Trumpet, 6566 preshape sphere, 910, 12, 14, 7879, 84, 133, 165 real projective space, 5559, 67, 76, 127129 Riemannian m., 4748, 50, 52, 60, 62, 78, 84, 88, 121 sphere as example of m., 38, 50, 5459, 66, 76, J23, 127 129, 131
sphere 'of preshapes, see preshape .sphere submanifold,52 tangent vector in a m., see Tangent vector topological m., 37, 39 torus as example. ofa m., 38, 55, 66 Mathematical statistics, 200 Matrix association m., 89 block, 196, 198 characteristic equation, see eigenvalue columns of am., 90, 99 covariance m., 130, 132133 determinant of m., 103, 122 diagonal m., 32 distance m., 8889, 91, 116 eigenvalue of m., 31, 8992, 102104 characteristic equation for e., 9798, 102103 e. as perturbation of unity, 102 moments of e., 102, 104105 eigenvector of m., 8990, 104 principal e., 92 Helmert m., 130131, 133, 165 Jacobian m., 36, 37, 53, 111, 113, 124, 127 main diagonal of m., 100 minors of m., 103 nonnegative definite, 89, 91 see also positive definite symmetric m. orthogonal, 3033, 100, 111, 113, 115, J27, 131 perturbation of identity m., 9799 pixel m., 195 positive definite symmetric m., 47 see also nonnegative definite presizeandshape m., see Presizeandshape matrix rank of m., 32 rows of m., .90 shape m., see Shape. matrix
singular value decomposition, 31~32, 95, 9798, 101, 104 sizeandshape m., see Sizeandshape matrix special orthogonal, 30 special unitary, 31 symmetric, 88, 90 trace of m., 8, 14, 103 unitary, 3031 upper triangular, 97, 100101, 115 Maximum internal angle, 27, 162 Mean of a sample, see Centroid Mean shape, 180 Mean vector, 130 Median direction, see Directional median see also Mt, Tom dinosaur tracks Medical sciences, 2, 199 Megalithic sites, 158 Method of coordinates, 24, 26 Metric, 12, 28,60 FubiniStudy metric, 62 equivalent to .Procrustean metric, 78 Metric space, 12, 34, 50 Metric tensor, 4749, 5152, 54, 5758, 60, 62, 84, 8687, 99, 102, 106, 124, 126, 147 and volume in manifolds, 122 invariance of m. t. under relabeling, 103104 invariance under right multiplication, 104, 115 m. t. for upper triangular shape representations, 101 as quadratic form on elements of dA, 102103 sundry examples, see Manifold Microscopy, 200 Minimum variance equivariant estimation, 3 Moebius transformation, see Transformations Mt. Tom dinosaur tracks, vi, 1620, 173182 bipedal, tridactylic species, 174 footprint classification, 1920, 173174
223
footprint condition, 173 species of dinosaurs, 1920 Anchisauripus, 19, 174, 178 Eubronies, 1920, 173174, 178179, 182 Grallator, 19, 174, 178 therapod, 174 trackway orientation, 1920, 174176 directional median, 175 histogram, 175 trackway scale analysis, 176178 boxplot of stride lengths, 177 footprint length, 176, 178 Froude numbers, 176 speed formula, 176, 178 stride length, 20, 176179 trackway shape analysis, 20, 178182 geodesic distance versus stride length, 179180 MardiaDryden density, 18~182
mean shape, 180 stretching effect, 179 uncertainty in classification, 174 Multidimensional scaling, 88 metric scaling, 88 nonmetric scaling, 88 see also Principal coordinate analysis Multivariate morphometries, 2, 6 Multivariate normal distribution, see Distribution, normal Multivariate statistics, 79, 200
Nearest neighbor, 139140, 190191 kth nearest neighbor, 139140 Nonsphericity property, 140, 142 . Norm on space of upper triangular matrices, 102 Normal distribution, see Distribution, normal
Obtuse angle in triangle, 171 Open set, 29, 37, 39, 52, 54, 5657, 118119
224
Index
Orbit, 1012, 59, 62, 165 Orbit space, 11 Orientation function, 11 Orientation information, 711, 84, 100 Orthogonal matrix, see Matrix Orthogonal transformation, see Transformations Orthogonality of vectors, 90 Orthonormal vectors, 99, 122
p'dimensional volume, 30 Paleontology, 199 Parabolic approximation to circular arc, 160 Parallelepiped, 122 Partial derivative, 36, 49, 51 Path in a manifold, 4245, 48, 53 geodesic, 4851, 57, 6265 great circle in sphere, 54, 5761 helix as geodesic in cylinder, 67 horizontal geodesic, 6162, 87 tangent paths in m., 4445, 53 Pathwise connected manifold, 50 Pattern, 184 Pattern recognition, 200 Permutation, 103·104 Pillow problems of Lewis Carroll, 171 Pixel value, 195 Poincare Disk, see Manifold Poincare Plane, see Manifold Poincare Trumpet, see Manifold Point at infinity, see Complex plane Point processes in manifold, 134145 binomial process, 134138, 143 of lines, 135137 locally finite, 138 Poisson process, 2, 134145, 168, 184, 1P9 homogenoous, 139 intensity of P. p., 138139 particles in P. p., 138 volumepreserving, 138139 Poisson approximation, 143, 148, 186, 188189
Index Poisson distribution, see Distribution Poisson process, see Point process Pontogram, 162 Post molds from Late Bronze Age England, 2024, 182184, 190193 Aldermaston Wharf, 20, 182, 190193 circle of post molds, 2324, 182184,191192 annular criterion, 23, 187190, 191192 expected number, 187, 189, 191192 radius, 23 clusters of post molds, 23, 182, 190191 interpoint distances, 20, 190191 post mold patterns, 20, 23 expected number, 185186 region of post mold activity, 187, 190191 roundhouses, 20, 23, 182184, 187, 191192 South Lodge Camp, 20, 23, 190193 Thorny Down, 182184, 187 Preshape sphere, see Manifold Preshape statistic, 914, 16, 58, 76, 79, 133134, 149 Presizeandshape matrix, 99100 Principal component analysis, 88, 91 Principal coordinate analysis, 8794 application to Iron Age brooches, see Brooches from Iron Age Probability distribution, see Distribution Probability measure, 4, 118, 123, 146 Probability space, 118, 123 Probability theory, v, 27, 119 Procrustean distance or metric, 3, 1314, 16, 28, 60, 72, 76, 9192, 167 equivalent to FubiniStudy metric, 78
matrix of interpoint Procrustean distances, 88 on general shape spaces, 80 Procrustean school of shape, see Kendall school of shape Procrustes analysis, 3, 6 Procrustes distance or metric, see Procrustean distance Psychometrics, 3
Quadratic equation, 98 Quiver diagram, 198
RadonNikodym derivative, 125 ratio of volume elements, 124125 Random quadrilateral, 27 Random set, 135 Random shape, 27, 139, 149 Random triangle, 4, 27 Random variable, see Statistic on a manifold Random vector, see Statistic on a manifold Real part of a complex number, 12, 31, 6970, 72, 160 Rectangle, 172 Rectangular lattice, 195 Residuals about centroid, 8 Riemannian manifold, see Manifold Riemannian metric, see Metric tensor Riemannian submersion, see FUnction Right triangle, 115 Rotation, see Transformations
Sample mean, see Centroid Sample space, 117118 Scale change, see Transformations Scale information, 711, 100 Scale parameters, 3 Secant vector, 43 Shape coordinates, 11, 27, 69 Bookstein coordinates, 6974, 77, 9799, 105, 150157
225
degeneracies when landmarks coincide, 71 generalized Bookstein coordinates, 100101, 105 on the sphere, 73 upper triangular shape representation, 101102, 114 Shape difference or variation, 24, 26, 35 Shape manifold, vi, 1,4, 1112, 14, 26, 28, 5859, 69, 72 Casson spheres, 81 proof that C. s. is topological sphere, 81 singularity set in C. s. and other shape manifolds, 8485 complex projective space of planar shapes, 7779, 88, 149150 geometry of E~ versus E~, 8182 hemisphere of triangle shapes in R 3,82 Kendall's shape spaces for landmarks in dimensions three and higher, 7987 Poincare half plane of triangle shapes, 2, 9599, 114 real projective space of shapes of onedimensional landmarks, 58 shape manifolds with boundary, 81 simplex shape spaces, 95106, 111, 114 singularities in shape manifolds, 81, 8384, 87 see also Shape manifold, Casson spheres sphere of triangle shapes, 1, 6977,81, 114115, 150, 165 great circle of collinear triangles, 74, 77 great circles of isosceles triangles, 74 Shape matrix, 100 Shape of line configuration, 137 Shape of triangles, 27, 6977 shape of collinear t., 7576
226
Index
Index
Shape of triangles (cont.) shape of equilateral t., 7274, 76, 155 shape of isosceles t., 74 Sigmafield,117119 sigmafield generated by class, 118,147 Similar sets, 35 Similar triangles, 6, 114115 Simplex, 30, 99101, 141 Simplex shape, 143 Simplex shape space, see Shape manifolds Singular value decomposition, see Matrix Singularities in shape manifolds, see Shape manifolds Sizeandshape matrix, 100 Size variable, 46, 27 Skull shapes and images, 1417,24, 107,113 Spatial interpolation, see Interpolation Special orthogonal transformation, see Transformations Special unitary transformation, see Transformations Sphere, see Manifold and Shape manifold Sphere of preshapes, see Manifold Spline, see Thinplate spline Standardization of data sets, 9 Statistic on a manifold, 118121 random variable, 119121, 130 random vector, 119121 Stereographic 'projection, see Transformations Stereology, 200 Stochastic geometry, 3, 199 Stochastic independence, 121 Straight line as example of geodesic, 51 Stretch factor, 160, 171 Submersion, see Function Subspace, 52, 7778, 84, 100, 115 Surface area on 2sphere, 123, 127 Symmetric function, 184
Tangent approximation to shape variation, 16, 170171 t. a. and concentration parameter, 171 Tangent paths in a manifold, see Path in a manifold Tangent vector, 38, 4248, 51, 62 basis vectors for the tangent space, 46, 5253, 57 length of tangent vector, 48, 50 orientation of tangent vector, 50 scalar multiplication of tangent vectors, 45,67 sum of tangent vectors, 45, 67 tangent space, 4243, 46, 5153, 62,84 horizontal tangent space, 8687 vertical tangent space, 8687 tangent vector field, 46, 66, 197 transporting vectors using affine connection, 50 Taylor approximation, 36 Tensor, see Metric tensor Tessellation, 141 Delaunay, see Delaunay tessellation Tetrahedral shapes, 105106 Thinplate splines, 106110 closed under similarity transformations, 110 landmarks as knots of the spline, 107 metal plate interpretation, 108 not bijective, 110 not invariant under function inversion, 110 see also Bending energy Topological singularity, 85 Topological space, 3739 Topology, 37, 39, 83 construction on general shape spaces, 80 Transformations, 106, 125 affine t., 2930, 64, 9597, 106, 111, 130, 152157 shearing effect of a. t., 96, 111 areapreserving, 157 conformal, 111112
diffeomorphism, 37, 39, 41, 44, 55, 6667, 113, 124125 Euclidean motion, 4, 3435, 135, 139 Helmert, 130 homeomorphism, 37, 39, 72 inversion, 112 isometries, 34, 5758, 64, 104, 125, 134 isometries of complex projective spaces, 129 isometries of pspheres, 127129 isometries of real projective spaces, 127129 isometries of the sphere of preshapes, 80 isometries of the sphere of triangle shapes, 73 isometry between 8 1 (1/2) and nr', 76 isometry between L;~ and 8 2 (1/ 2), 76, 115 isometry between L;2 and cpn2, 78 local isometry, 58, 113 see also linear isometry isotropic rescaling or scale .change, see scale transformation Jacobian of t., 37, 112, 124, 144 linear fractional t., see Moebius transformation linear isometry, 34, 57, 78, 86 linear t., 2930, 32, 36, 53, 57, 105, 153157 Moebius t., 7273, 112 orientationpreserving t., 111 orthogonal t., 3033, 59, 100, 127129, 147 reflection, 31, 34, 58, 76, 8182, 101 rescaling, see scale transformation rotation, v, 3, 10, 30, 58, 70, 72, 77, 79,8182, 8485, 101, 157
227
scale t., v , 3, 9, 3435, 70, 77, 113 shapepreserving t., see similarity transformation similarity t., 3, 3435, 69, 77, 82, 95, 97, 101, 110113, 115, 137 special orthogonal t., 30, 3334, 7980, 84 special unitary t., 31 stereographic projection, 7174, 7677 translation, v, 3, 34, 70, 137 unitary t., 3033,59, 127, 129, 147 volumepreserving t., 112 Transformations of statistics, 124125, 144, 150, 152157 Translate of a set, 188 Translation, see Transformations Triangle, 35 Triangle inequality, 60 Trigonometric series, 157 Undirected line, 137, 148 Unit circle, 10, 13, 16, 31, 55, 76 Unitary matrix, see Matrix Unitary transformation, see Transformations Upper triangular matrix, see Matrix Upper triangular shape representation, see Shape coordinates
Vector of residuals, 131 Vector sum, see Tangent vector Vertical tangent space, see Tangent vector, tangent space Volume element, dVp 122, 131 Volume in a manifold, 121123 Von Neumann norm, see Norm on the space of upper triangular matrices Voronoi tessellation, 146 duality with Delaunay tessellation, 146
Springer Series in Statistics (continued from p. UJ
Pollard: Convergence of Stochastic Processes. Pratt/Gibbons: Concepts of Nonparametric Theory. Read/Cressie: GoodnessofFit Statistics for Discrete Multivariate Data. Reinsel: Elements of Multivariate Time Series Analysis. Reiss: A Course on Point Processes. Reiss: Approximate Distributions of Order Statistics: With Applications
to Nonparametric Statistics. Rieder: Robust Asymptotic Statistics. Rosenbaum: Observational Studies. Ross: Nonlinear Estimation. Sachs: Applied Statistics: A Handbook of Techniques, 2nd edition. SdrndaltSwenssoniWretman: Model Assisted Survey Sampling. Schervish: Theory of Statistics. Seneta: NonNegative Matrices and Markov Chains, 2nd edition. Shao/Tu: The Jackknife and Bootstrap. Siegmund: Sequential Analysis: Tests and Confidence Intervals. Simonoff: Smoothing Methods in Statistics. Small: The Statistical Theory of Shape. Tanner: Tools for Statistical Inference: Methods for the Exploration of Posterior
Distributions and Likelihood Functions, 3rd edition. Tong: The Multivariate Normal Distribution. van der Vaart/Wellner: Weak Convergence and Empirical Processes: With
Applications to Statistics. Vapnik: Estimation of Dependences Based on Empirical Data. Weerahandi: Exact Statistical Methods for Data Analysis. West/Harrison: Bayesian Forecasting and Dynamic Models. Wolter: Introduction to Variance Estimation. Yaglom: Correlation Theory of Stationary and Related Random Functions I:
Basic Results. Yaglom: Correlation Theory of Stationary and Related Random Functions II:
Supplementary Notes and References.